Signal Processing Methods

for Music Transcription

Signal Processing Methods
for Music Transcription

With 124 illustrations

Anssi Klapuri
Manuel Davy
Editors

Springer
 Manuel Davy
Anssi Klapuri
LAGIS/CNRS
Tampere University of Technology
Ecole Centrale de Lille
Institute of Signal Processing
Cite Scientifique
Korkeakoulunkatu 1
BP48
33720 Tampere, Finland
59651 Villeneuve d'Ascq
Anssi.Klapuritut.fi
Cedex, France
[email protected]

Cover design by Alex Gerasev.

Mathematics Subject Classification (2000): 68T10, 62F15, 68P20

Library of Congress Control Number: 2006920280

ISBN-10: 0-387-30667-6 e-ISBN: 0-387-32845-9

ISBN-13: 978-387-30667-4

Printed on acid-free paper.

2006 Springer Science-|-Business Media LLC

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science-I-Business Media LLC, 233 Spring Street, New York,
NY 10013, USA) and the author, except for brief excerpts in connection with reviews or scholarly
analysis. Use in connection with any form of information storage and retrieval, electronic adaptation,
computer software, or by similar or dissimilar methodology now known or hereafter developed is
forbidden.
The use in this publication of trade names, trademarks, service marks and similar terms, even if they
are not identified as such, is not to be taken as an expression of opinion as to whether or not they are
subject to proprietary rights.
Printed in the United States of America. (KeS/MP)

987654321

springer.com
Contents

Preface ix

List of Contributors xi

Part I Foundations

1 Introduction to Music Transcription

Anssi Klapuri 3
1.1 Terminology and Concepts 7
1.2 Perspectives on Music Transcription 11
1.3 Outline 17

2 A n Introduction to Statistical Signal Processing and

Spectrum Estimation
Manuel Davy 21
2.1 Frequency, Time-Frequency, and Cepstral Representations 21
2.2 Basic Statistical Methods 28
2.3 Bayesian Statistical Methods 39
2.4 Pattern Recognition Methods 52

3 Sparse Adaptive Representations for Musical Signals

Laurent Daudet, Bruno Torresani 65
3.1 Introduction 65
3.2 Parametric Representations 68
3.3 Waveform Representations 70
3.4 Conclusion 97
 Contents

Part II Rhythm and Timbre Analysis

4 Beat Tracking and Musical Metre Analysis

Stephen Hainsworth 101
4.1 Introduction 101
4.2 Summary of Beat-Tracking Approaches 102
4.3 Musical Background to Rhythmic Structure 105
4.4 Onset Detection 107
4.5 Rule-Based Approaches Ill
4.6 Autocorrelation Methods 112
4.7 Oscillating Filter Approaches 113
4.8 Histogramming Methods 115
4.9 Multiple Agent Approaches 116
4.10 Probabilistic Models 117
4.11 Comparison of Algorithms 124
4.12 Conclusions 127

5 Unpitched Percussion Transcription

Derry FitzGerald, Jouni Paulus 131
5.1 Introduction 131
5.2 Pattern Recognition Approaches 133
5.3 Separation-Based Approaches 142
5.4 Musicological Modelling 153
5.5 Conclusions 160
5.6 Acknowledgements 162

6 Automatic Classification of Pitched Musical Instrument

Sounds
Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy 163
6.1 Introduction 163
6.2 Methodology 169
6.3 Features and Their Selection 171
6.4 Classification Techniques 184
6.5 Classification of Isolated Sounds 188
^.^ Classification of Sounds from Music Files 193
6.7 Conclusions 199

Part III Multiple Fundamental Frequency Analysis

7 Multiple Fundamental Frequency Estimation Based on

Generative Models
Manuel Davy 203
7.1 Noisy Sum-of-Sines Models 204
7.2 Off-line Approaches 210
 Contents vii

7.3 On-Line Approaches 217

7.4 Other On-Line Bayesian Approaches 225
7.5 Conclusions 227

8 Auditory Model-Based Methods for Multiple Fundamental

Frequency Estimation
Anssi Klapuri 229
8.1 Introduction 229
8.2 Musical Sounds and FO Estimation 231
8.3 Pitch Perception Models 234
8.4 Using an Auditory Model as a Front End 244
8.5 Computational Multiple FO Estimation Methods 248
8.6 Conclusions 264

9 Unsupervised Learning Methods for Source Separation in

Monaural Music Signals
Tuomas Virtanen 267
9.1 Introduction 267
9.2 Signal Model 268
9.3 Independent Component Analysis 274
9.4 Sparse Coding 278
9.5 Non-Negative Matrix Factorization 282
9.6 Prior Information about Sources 284
9.7 Further Processing of the Components 286
9.8 Time-Varying Components 289
9.9 Evaluation of the Separation Quality 294
9.10 Summary and Discussion 295

Part IV Entire Systems, Acoustic and Musicological Modelling

10 Auditory Scene Analysis in Music Signals

Kunio Kashino 299
10.1 Introduction 299
10.2 Strategy for Music Scene Analysis 304
10.3 Probabilistic Models for Music Scene Analysis 313
10.4 Conclusion: From Grouping to Generative Estimation 324
11 Music Scene Description
Masataka Goto 327
11.1 Introduction 327
11.2 Estimating Melody and Bass Lines 330
11.3 Estimating Beat Structure 341
11.4 Estimating Drums 342
11.5 Estimating Chorus Sections and Repeated Sections 342
viii Contents

11.6 Evaluation Issues 355

11.7 Applications of Music Scene Description 355
11.8 Conclusion 358

12 Singing Transcription
Matti Ryyndnen ! 361
12.1 Introduction 361
12.2 Singing Signals 364
12.3 Feature Extraction 368
12.4 Converting Features into Note Sequences 375
12.5 Summary and Discussion 390

References 391

Index 429
Preface

Signal processing techniques, and information technology in general, have un-

dergone several scientific advances which permit us to address the very com-
plex problem of automatic music transcription (AMT). During the last ten
years, the interest in AMT has increased rapidly, and the time has come for
a book-length overview of this subject.
The purpose of this book is to present signal processing algorithms ded-
icated to the various aspects of music transcription. AMT is a multifaceted
problem, comprising several subtasks: rhythm analysis, multiple fundamental
frequency analysis, sound source separation, musical instrument classification,
and integration of all these into entire systems. AMT is, in addition, deeply
rooted in fundamental signal processing, which this book also covers. As the
field is quite wide, we have focused mainly on signal processing methods and
Western polyphonic music. An extensive presentation of the work in musicol-
ogy and music perception is beyond the scope of this book.
This book is mainly intended for researchers and graduate students in sig-
nal processing, computer science, acoustics, and music. We hope that the book
will make the field easier to approach, providing a good starting point for new-
comers, but also a comprehensive reference source for those already working in
the field. The book is also suitable for use as a textbook for advanced courses
in music signal processing. The chapters are mostly self-contained, and read-
ers may want to read them in any order or jump from one to another at will.
Whenever an element from another chapter is needed, an explicit reference is
made to the relevant chapter. Chapters 1 and 2 provide some background of
AMT and signal processing for the entire book, respectively. Otherwise, only
a basic knowledge of signal processing is assumed.
Editing a book is a great deal of work. This volume was made possible
by those who provided us support and help. We would like to thank VaishaU
Damle and Ana Bozicevic at Springer for their help and support, and for
their quick replies to our e-mails. Also thanks to Teemu Karjalainen for his
practical assistance with WT^K^.
X Preface

Early versions of individual chapters were reviewed by the following peo-

ple, whose valuable comments and suggestions are gratefully acknowledged:
- Michael Casey, Goldsmiths College, University of London, UK
- A. Taylan Cemgil, University of Cambridge, UK
- Alain de Cheveigne, Ecole Normale Superieure, Paris, Prance
- Simon Dixon, Austrian Research Institute for Artificial Intelligence, Vienna
- Dan Ellis, Columbia University, New York
- Olivier Gillet, Telecom-Paris (ENST), Paris, France
- Aki Harma, Philips Research Laboratories, Eindhoven, The Netherlands
- Marc Leman, Ghent University, Belgium
- Emanuele Pollastri, Erazero S.r.l., Milan, Italy

Thanks go also to the chapter authors, many of whom proofread another,

related chapter in the book and provided helpful comments.

Tampere, Finland Anssi Klapuri

Lille, France Manuel Davy
December 2005
List of Contributors

Laurent Daudet Perfecto Herrera-Boyer

Laboratoire d'Acoustique Musicale Institut Universitari de 1'Audiovisual
11 rue de Lourmel Universitat Pompeu Fabra
75015 Paris, Prance Pg. CircumvaMacio 8
daudetQlam.jussieu.fr 08003 Barcelona, Spain
Manuel Davy pherreraQiua.upf.es
LAGIS/CNRS
BP 48, Cite Scientifique Kunio Kashino
59651 Villeneuve d'Ascq Cedex NTT Communication Science
Prance Laboratories
[email protected] Nippon Telegraph and Telephone
Derry FitzGerald Corporation 3-1
Cork Institute of Technology Morinosato-Wakamiya
Rossa Avenue Atsugi, 243-0198, Japan
Bishopstown kimioOeye.brl.ntt.co.jp
Cork, Ireland
derry.fitzgeraldQcit.ie Anssi Klapuri
Masataka Goto Institute of Signal Processing
National Institute of Advanced Tampere University of Technology
Industrial Science and Technology Korkeakoulunkatu 1
(AIST) 33720 Tampere, Finland
1-1-1 Umezono, Tsukuba Anssi.KlapuriOtut.fi
Ibaraki 305-8568, Japan
m.gotoQaist.go.jp Jouni Paulus
Stephen Hainsworth Institute of Signal Processing
Tillinghast-Towers Perrin Tampere University of Technology
71 High Holborn Korkeakoulunkatu 1
London WCIV 6TH, UK 33720 Tampere, Finland
swh210cantab.net Jouni.PaulusQtut.fi
 List of Contributors

Bruno Torresani Korkeakoulunkatu 1

Labor at oire d'Analyse 33720 Tampere, Finland
Topologie et Probabilites Matti.RyynanenOtut.fi
CMI, Universite de Provence
39 rue F. Joliot-Curie
Tuomas Virtanen
13453 Marseille cedex 13, France
Institute of Signal Processing
Bruno.TorresaniQcmi.univ-mrs.fr
Tampere University of Technology
Matti Ryynanen Korkeakoulunkatu 1
Institute of Signal Processing 33720 Tampere, Finland
Tampere University of Technology Tuomas.VirtanenOtut.fi
 Part I

Foundations
Introduction to Music Transcription

Anssi Klapuri

Institute of Signal Processing, Tampere University of Technology

Korkeakoulunkatu 1, 33720 Tampere, Finland
Anssi.KlapuriQtut.fi

Music transcription refers to the analysis of an acoustic musical signal so as

to write down the pitch, onset time, duration, and source of each sound that
occurs in it. In Western tradition, written music uses note symbols to indicate
these parameters in a piece of music. Figures 1.1 and 1.2 show the notation of
an example music signal. Omitting the details, the main conventions are that
time flows from left to right and the pitch of the notes is indicated by their
vertical position on the staff lines. In the case of drums and percussions, the
vertical position indicates the instrument and the stroke type. The loudness
(and the applied instrument in the case of pitched instruments) is normally
not specified for individual notes but is determined for larger parts.
Besides the common musical notation, the transcription can take many
other forms, too. For example, a guitar player may find it convenient to read
chord symbols which characterize the note combinations to be played in a
more general manner. In a computational transcription system, a MIDI file^
is often an appropriate format for musical notations (Fig. 1.3). Common to
all these representations is that they capture musically meaningful parameters
that can be used in performing or synthesizing the piece of music in question.
From this point of view, music transcription can be seen as discovering the
'recipe', or reverse-engineering the 'source code' of a music signal.
A complete transcription would require that the pitch, timing, and instru-
ment of all the sound events be resolved. As this can be very hard or even
theoretically impossible in some cases, the goal is usually redefined as being
either to notate as many of the constituent sounds as possible (complete tran-
scription) or to transcribe only some well-defined part of the music signal, for
example the dominant melody or the most prominent drum sounds (partial
transcription). Both of these goals are relevant and are discussed in this book.
Music transcription is closely related to structured audio coding. A musical
notation or a MIDI file is an extremely compact representation that retains

^Musical Instrument Digital Interface (MIDI) is a standard for exchanging per-

formance data and parameters between electronic musical devices [462], [571].
 Anssi Klapuri

t 0-1

Fig. 1.1. An acoustic musical signal (top) and its time-frequency domain represen-
tation (bottom). The excerpt is from Song G034 in the RWC database [230].

Fig. 1.2. Musical notation corresponding to the signal in Fig. 1.1. The upper staff
lines show the notation for pitched musical instruments and the lower staff lines
show the notation for percussion instruments.

C4

.
~ i
- C3
_

1= CI

^^ _ . ^ _ i

Fig. 1.3. A 'piano-roll' illustration of a MIDI file which corresponds to the pitched
instruments in the signal in Fig. 1.1. Different notes are arranged on the vertical
axis and time flows from left to right.
 1 Introduction 5

the characteristics of a piece of music to an important degree. Another related

area of study is that of music perception [144]. Detecting and recognizing
individual sounds in music is a big part of its perception, although it should
be emphasized that musical notation is primarily designed to serve sound
production and not to model hearing. We do not hear music in terms of note
symbols but, as described by Bregman [49, pp. 457-460], music often 'fools'
the auditory system so that we perceive simultaneous sounds as a single entity.
In addition to audio coding, applications of music transcription comprise
Music information retrieval based on the melody of a piece, for example.
Music processing^ such as changing the instrumentation, arrangement, or
the loudness of different parts before resynthesizing a piece from its score.
Human-computer interaction in various applications, including score type-
setting programs and musically oriented computer games. Singing tran-
scription is of particular importance here.
Music-related equipment, ranging from music-synchronous light effects to
highly sophisticated interactive music systems which generate an accom-
paniment for a soloist.
Musicological analysis of improvised and ethnic music for which musical
notations do not exist.
Transcription tools for amateur musicians who wish to play along with
their favorite music.
The purpose of this book is to describe algorithms and models for the dif-
ferent subtopics of music transcription, including pitch analysis, metre analysis
(see Section 1.1 for term definitions), percussion transcription, musical instru-
ment classification, and music structure analysis. The main emphasis is laid
on the low-level signal analysis where sound events are detected and their
parameters are estimated, and not so much on the subsequent processing of
the note data to obtain larger musical structures. The theoretical background
of different signal analysis methods is presented and their application to the
transcription problem is discussed.
The primary target material considered in this book is complex music
signals where several sounds are played simultaneously. These are referred
to as polyphonic signals, in contrast to monophonic signals where at most
one note is sounding at a time. For practical reasons, the scope is limited to
Western music, although not to any particular genre. Many of the analysis
methods make no assumptions about the larger-scale structure of the signal
and are thus applicable to the analysis of music from other cultures as well.
To give a reasonable estimate of the achievable goals in automatic music
transcription, it is instructive to study what human listeners are able to do in
this task. An average listener perceives a lot of musically relevant information
in complex audio signals. He or she can tap along with the rhythm, hum the
melody (more or less correctly), recognize musical instruments, and locate
structural parts of the piece, such as the chorus and the verse in popular music.
Harmonic changes and various details are perceived less consciously. Similarly
6 Anssi Klapuri

to natural language, however, reading and writing music requires education.

Not only the notation needs to be studied, but recognizing different pitch
intervals and timing relationships is an ability that has to be learned - these
have to be encoded into a symbolic form in one's mind before writing them
down. Moreover, an untrained listener is typically not able to hear the inner
lines in music (sub-melodies other than the dominant one), so musical ear
training is needed to develop an analytic mode of listening where these can be
distinguished. The richer the polyphonic complexity of a musical composition,
the more its transcription requires musical ear training and knowledge of
the particular musical style and of the playing techniques of the instruments
involved.
First attempts towards the automatic transcription of polyphonic music
were made in the 1970s, when Moorer proposed a system for transcribing two-
voice compositions [477], [478]. His work was followed by that of Chafe et al.
[81], Piszczalski [520], and Maher [426], [427] in the 1980s. In all these early
systems, the number of concurrent voices was limited to two and the pitch
relationships of simultaneous sounds were restricted in various ways. On the
rhythm analysis side, the first algorithm for beat tracking^ in general au-
dio signals was proposed by Goto and Muraoka in the 1990s [235], although
this was preceded by a considerable amount of work for tracking the beat
in parametric note data (see [398] for a summary) and by the beat-tracking
algorithm of Schloss for percussive audio tracks [567]. First attempts to tran-
scribe percussive instruments were made in the mid-1980s by Schloss [567] and
later by Bilmes [37], both of whom classified different types of conga strikes
in continuous recordings. Transcription of polyphonic percussion tracks was
later addressed by Goto and Muraoka [236]. A more extensive description of
the early stages of music transcription has been given by Tanguiane in [619,
pp. 3-6].
Since the beginning of 1990s, the interest in music transcription has grown
rapidly and it is not possible to make a complete account of the work here.
However, certain general trends and successful approaches can be discerned.
One of these has been the use of statistical methods. To mention a few exam-
ples, Kashino [332], Goto [223], Davy and Godsill [122], and Ryynanen [559]
proposed statistical methods for the pitch analysis of polyphonic music; in
beat tracking, statistical methods were employed by Cemgil and Kappen [77],
Hainsworth and MacLeod [266], and Klapuri et al. [349]; and in percussive in-
strument transcription by Gillet and Richard [209] and Paulus et al. [506]. In
musical instrument classification, statistical pattern recognition methods pre-
vail [286]. Another trend has been the increasing utilization of computational
models of the human auditory system. These were first used for music tran-
scription by Martin [439], and auditorily motivated methods have since been
proposed for polyphonic pitch analysis by Karjalainen and Tolonen [627] and

^ Beat tracking refers to the estimation of a rhythmic pulse which corresponds to

the tempo of a piece and (loosely) to the foot-tapping rate of human listeners.
 1 Introduction 7

Klapuri [354], and for beat tracking by Scheirer [564], for example. Another
prominent approach has been to model the human auditory scene analysis
(ASA) ability. The term ASA refers to the way in which humans organize
spectral components to their respective sounds sources and recognize simul-
taneously occurring sounds [49]. The principles of ASA were brought to the
pitch analysis of polyphonic music signals by Mellinger [460] and Kashino
[333], and later by Godsmark and Brown [215] and Sterian [609]. Most re-
cently, several unsupervised learning methods have been proposed where a
minimal number of prior assumptions are made about the analysed signal.
Methods based on independent component analysis [304] were introduced to
music transcription by Casey [70], [73], and various other methods were later
proposed by Lepain [403], Smaragdis [598], [600], Abdallah [2], [5], Virtanen
(see Chapter 9), FitzGerald [186], [188], and Paulus [505]. Of course, there
are also methods that do not represent any of the above-mentioned trends,
and a more comprehensive review of the literature is presented in the coming
chapters.
The state-of-the-art music transcription systems are still clearly inferior to
skilled human musicians in accuracy and flexibility. That is, a reliable general-
purpose transcription system does not exist at the present time. However,
some degree of success has been achieved for polyphonic music of limited
complexity. In the transcription of pitched instruments, typical restrictions
are that the number of concurrent sounds is limited [627], [122], interference of
drums and percussive sounds is not allowed [324], or only a specific instrument
is considered [434]. Some promising results for the transcription of real-world
music on CD recordings has been demonstrated by Goto [223] and Ryynanen
and Klapuri [559]. In percussion transcription, quite good accuracy has been
achieved in the transcription of percussive tracks which comprise a limited
number of instruments (typically bass drum, snare, and hi-hat) and no pitched
instruments [209], [505]. Also promising results have been reported for the
transcription of the bass and snare drums on real-world recordings, but this
is a more open problem (see e.g. Zils et al. [693], FitzGerald et al. [189],
Yoshii et al. [683]). Beat tracking of complex real-world audio signals can
be performed quite reliably with the state-of-the-art methods, but difficulties
remain especially in the analysis of classical music and rhythmically complex
material. Comparative evaluations of beat-tracking systems can be found in
[266], [349], [248]. Research on musical instrument classification has mostly
concentrated on working with isolated sounds, although more recently this
has been attempted in polyphonic audio signals, too [331], [33], [170], [647].

1.1 Terminology and Concepts

Before turning to a more general discussion of the music transcription problem

and the contents of this book, it is necessary to introduce some basic terminol-
ogy of auditory perception and music. To discuss music signals, we first have
8 Anssi Klapuri

to discuss the perceptual attributes of sounds of which they consist. There are
four subjective quahties that are particularly useful in characterizing sound
events: pitch, loudness, duration, and timbre [550].
Pitch is a perceptual attribute which allows the ordering of sounds on
a frequency-related scale extending from low to high. More exactly, pitch is
defined as the frequency of a sine wave that is matched to the target sound
by human listeners [275]. Fundamental frequency (FO) is the corresponding
physical term and is defined for periodic or nearly periodic sounds only. For
these classes of sounds, FO is defined as the inverse of the period and is closely
related to pitch. In ambiguous situations, the period corresponding to the
perceived pitch is chosen.
The perceived loudness of an acoustic signal has a non-trivial connection
to its physical properties, and computational models of loudness perception
constitute a fundamental part of psychoacoustics [523].^ In music processing,
however, it is often more convenient to express the level of sounds with their
mean-square power and to apply a logarithmic (decibel) scale to deal with the
wide dynamic range involved. The perceived duration of a sound has more or
less one-to-one mapping to its physical duration in cases where this can be
unambiguously determined.
Timbre is sometimes referred to as sound 'colour' and is closely related to
the recognition of sound sources [271]. For example, the sounds of the violin
and the fiute may be identical in their pitch, loudness, and duration, but are
still easily distinguished by their timbre. The concept is not explained by any
simple acoustic property but depends mainly on the coarse spectral energy
distribution of a sound, and the time evolution of this. Whereas pitch, loud-
ness, and duration can be quite naturally encoded into a single scalar value,
timbre is essentially a multidimensional concept and is typically represented
with a feature vector in musical signal analysis tasks.
Musical information is generally encoded into the relationships between
individual sound events and between larger entities composed of these. Pitch
relationships are utihzed to make up melodies and chords. Timbre and loud-
ness relationships are used to create musical form especially in percussive
music, where pitched musical instruments are not necessarily employed at
all. Inter-onset interval (lOI) relationships, in turn, largely define the rhyth-
mic characteristics of a melody or a percussive sound sequence (the term lOI
refers to the time interval between the beginnings of two sound events). Al-
though durations of the sounds play a role too, the lOIs are more crucial in
determining the perceived rhythm [93]. Indeed, many rhythmically important
instruments, such as drums and percussions, produce exponentially decaying
wave shapes that do not even have a uniquely defined duration. In the case of

^Psychoacoustics is the science that deals with the perception of sound. In a

psychoacoustic experiment, the relationships between an acoustic stimulus and the
resulting subjective sensation are studied by presenting specific teisks or questions
to human listeners [550].
 1 Introduction
F# G# A#

cToL F G A B
! ! ! ! ! ! ! ! ! !
Fig. 1.4. Illustration of the piano keyboard (only three octaves are shown here).

sustained musical sounds, however, the durations are used to control articu-
lation. The two extremes here are 'staccato', where notes are cut very short,
and 'legato', where no perceptible gaps are left between successive notes.
A melody is a series of pitched sounds with musically meaningful pitch
and lOI relationships. In written music, this corresponds to a sequence of
single notes. A chord is a combination of two or more simultaneous notes.
A chord can be harmonious or dissonant, subjective attributes related to the
specific relationships between the component pitches and their overtone par-
tials. Harmony refers to the part of music theory which studies the formation
and relationships of chords.
Western music arranges notes on a quantized logarithmic scale, with 12
notes in each octave range. The nominal fundamental frequency of note n can
be calculated as 440 Hz x 2^/^^, where 440 Hz is an agreed-upon anchor point
for the tuning and n varies from 48 to 39 on a standard piano keyboard,
for example. According to a musical convention, the notes in each octave are
lettered as C, C # , D, D # , E, F, . . . (see Fig. 1.4) and the octave is indicated
with a number following this, for example A4 and A3 referring to the notes
with fundamental frequencies 440 Hz and 220 Hz, respectively.
There are of course instruments which produce arbitrary pitch values and
not just discrete notes like the piano. When playing the violin or singing, for
example, both intentional and unintentional deviations take place from the
nominal note pitches. In order to write down the music in a symbolic form,
it is necessary to perform quantization^ or perceptual categorization [60]: a
track of pitch values is segmented into notes with discrete pitch labels, note
timings are quantized to quarter notes, whole notes, and so forth, and timbral
information is 'quantized' by naming the sound sources involved. In some
cases this is not necessary but a parametric or semi-symbolic^ representation
suffices.
An important property of basically all musical cultures is that correspond-
ing notes in different octaves are perceived as having a special kind of sim-
ilarity, independent of their separation in frequency. The notes C3, C4, and
C5, for example, play largely the same harmonic role although they are not
interchangeable in a melody. Therefore the set of all notes can be described
as representing only 12 pitch classes. An individual musical piece usually re-
cruits only a subset of the 12 pitch classes, depending on the musical key of

*In a MIDI file, for example, the time values are not quantized.
10 Anssi Klapuri

4iit 111^(II' |iih'ft4

latum
Tactus (beat);
Bar lines:
174 175 176 177 178 179 180
Time (s)

Fig. 1.5. A music signal with three metrical levels illustrated.

the piece. For example, a piece in the C major key tends to employ the white
keys of the piano, whereas a piece in B major typically employs all the black
keys but only two white keys in each octave. Usually there are seven pitch
classes that 'belong' to a given key. These are called scale tones and they pos-
sess a varying degree of importance or stability in the key context. The most
important is the tonic note (for example C in the C major key) and often a
musical piece starts or ends on the tonic. Perception of pitch along musical
scales and in relation to the musical key of the piece is characteristic to tonal
music, to which most of Western music belongs [377].
The term musical metre has to do with the rhythmic aspects of music: it
refers to the regular pattern of strong and weak beats in a piece. Perceiving the
metre consists of detecting moments of musical emphasis in an acoustic signal
and filtering them so that the underlying periodicities are discovered [404],
[93]. The perceived periodicities, pulses, at different time scales (or levels)
together constitute the metre, as illustrated in Fig. 1.5. Perceptually the most
salient metrical level is the tactus, which is often referred to as the foot-tapping
rate or the beat The tactus can be viewed as the temporal 'backbone' of a piece
of music, making beat tracking an important subtask of music transcription.
Further metrical analysis aims at identifying the other pulse levels, the periods
of which are generally integer multiples or submultiples of the tactus pulse.
For example, detecting the musical measure pulse consists of determining the
number of tactus beats that elapse within one musical measure (usually 2
to 8) and aligning the boundaries of the musical measures (bar lines) to the
music signal.
Another element of musical rhythms is grouping, which refers to the way in
which individual sounds are perceived as being grouped into melodic phrases;
these are further grouped into larger musical entities in a hierarchical manner
[404]. Important to the rhythmic characteristics of a piece of music is how
these groups are aligned in time with respect to the metrical system.
The structure of a musical work refers to the way in which it can be sub-
divided into parts and sections at the largest time-scale. In popular music,
for example, it is usually possible to identify parts that we label as the cho-
rus, the verse, an introductory section, and so forth. Structural parts can be
detected by finding relatively long repeated pitch structures or by observing
considerable changes in the instrumentation at section boundaries.
 1 Introduction 11

The forthcoming chapters of this book address the extraction and analysis
of the above elements in musical audio signals. Fundamental frequency esti-
mation is considered in Parts III and IV of this book, with a separate treatise
on melody transcription in Chapters 11 and 12. Metre analysis is discussed
in Chapter 4 and percussion transcription in Chapter 5. Chapter 6 discusses
the measurement of timbre and musical instrument classification. Structure
analysis is addressed in Chapter 11, and the quantization of time and pitch
in Chapters 4 and 12, respectively. Before going to a more detailed outline
of each chapter, however, let us have a look at some general aspects of the
transcription problem.

1.2 Perspectives on Music Transcription

When starting to design a transcription system, certain decisions have to be
made before the actual algorithm development. Among the questions involved
are: How should the transcription system be structured into smaller submod-
ules or tasks? What kind of data representations would be the most suitable?
Should musical information be used as an aid in the analysis? Would it be
advantageous to analyse larger musical structures before going into note-by-
note transcription? These general and quite 'philosophical' issues are discussed
from various perspectives in the following.

1.2.1 Neurophysiological P e r s p e c t i v e

First, let us consider a neurophysiological argument about how the music

transcription problem should be decomposed into smaller subtasks. In human
auditory cognition, modularity of a certain kind has been observed, mean-
ing that certain parts can be functionally and neuro-anatomically isolated
from the rest [517], [687}^ [623]. One source of evidence for this are studies
with brain-damaged patients: an accidental brain damage may selectively af-
fect musical abilities but not speech-related abilities, and vice versa [518].
Moreover, there are patients who suffer from difficulties dealing with pitch
variations in music but not with temporal variations. In music performance
or in perception, either of the two can be selectively lost [29], [517].
Peretz has studied brain-damaged patients who suffer from specific music
impairments and she proposes that the music cognition system comprises at
least four discernable 'modules' [517], [518]. An acoustic analysis module seg-
regates a mixture signal into distinct sound sources and extracts the percep-
tual parameters of these (including pitch) in some raw form. This is followed
by two parallel modules which carry out pitch organization (melodic contour
analysis and tonal encoding of pitch) and temporal organization (rhythm and
metre analysis). The fourth module, musical lexicon, contains representations
of the musical phrases a subject has previously heard.
12 Anssi Klapuri

Neuroimaging experiments in healthy subjects are another way of local-

izing the cognitive functions in the brain. Speech sounds and higher-level
speech information are known to be preferentially processed in the left audi-
tory cortex, whereas musical sounds are preferentially processed in the right
auditory cortex [623]. Interestingly, however, when musical tasks specifically
involve processing of temporal information (temporal synchrony or duration),
the processing is weighted towards the left hemisphere [687], [517]. The rel-
ative (not complete) asymmetry between the two hemispheres seems to be
related to the acoustic characteristics of the signals: rapid temporal informa-
tion is characteristic for speech, whereas accurate processing of spectral and
pitch information is more important in music [581], [687], [623]. Zatorre et al.
proposed that the left auditory cortex is relatively specialized to a better time
resolution and the right auditory cortex to a better frequency resolution [687].
In computational transcription systems, rhythm and pitch have often been
analysed separately and using different data representations (see e.g. [332],
[439], [237], [223], [350], [122]). Typically, a better time resolution is applied
in rhythm analysis and a better frequency resolution in pitch analysis. Based
on the above studies, this seems to be justified to some extent. However, it
should be kept in mind that studying the human brain is very difficult and the
reported results are therefore a subject of controversy. Also, the structure of
transcription systems is often determined by merely pragmatic considerations.
For example, temporal segmentation is performed prior to pitch analysis in
order to allow an appropriate positioning of analysis frames in pitch analysis,
which is typically the most demanding stage computationally.

1.2.2 Human Transcription

Another viewpoint to the transcription problem is obtained by studying the

conscious transcription process of human musicians and by inquiring about
their transcription strategies. The aim of this is to determine the sequence of
actions or processing steps that leads to the transcription result.
As already mentioned above, reading and writing music is an acquired
ability and therefore the practice of music transcription is of course affected by
its teaching at musical institutions. In this context, the term musical dictation
is used to refer to an exercise where a musical excerpt is played and it has to
be written down as notes [206]. An excellent study on the practice of musical
dictation and ear training pedagogy can be found in [280].
Characteristic to ear training is that the emphasis is not on trying to hear
more but to recognize what is being heard; to hear relationships accurately
and with understanding. Students are presented with different pitch intervals,
rhythms, and chords, and they are trained to name these. Simple examples
are first presented in isolation and when these become familiar, increasingly
complex material is considered. Melodies are typically viewed as a synthesis of
pitch and rhythm. For example, Ghezzo instructs the student first to memorize
the fragment of music that is to be written down, then to write the pitch of
 1 Introduction 13

the notes, and finally to apply the rhythm [206, p. 6]. Obviously, ear training
presumes a normally hearing subject who is able to detect distinct sounds and
their pitch and timing in the played excerptsaspects which are very difficult
to model computationally.
Recently, Hainsworth conducted a study where he asked trained musicians
to describe how they transcribe realistic musical material [263]. The subjects
(19 in total) had transcribed music from various genres and with varying goals,
but Hainsworth reports that a consistent pattern emerged in the responses.
Most musicians first write down the structure of the piece, possibly with some
key phrases marked in an approximate way. Next, the chords of the piece or
the bass line are notated, and this is followed by the melody. As the last step,
the inner lines are studied. Many reported that they heard these by repeated
listening, by using an instrument as an aid, or by making musically educated
guesses based on the context.
Hainsworth points out certain characteristics of the above-described pro-
cess. First, it is sequential rather than concurrent; quoting the author, 'no-
one transcribes anything but the most simple music in a single pass'. In this
respect, the process differs from most computational transcription systems.
Secondly, the process relies on the human ability to attend to certain parts
of a polyphonic signal while selectively ignoring others.^ Thirdly, some early
analysis steps appear to be so trivial for humans that they are not even men-
tioned. Among these are style detection (causing prior expectations regarding
the content), instrument identification, and beat tracking.

1.2.3 Mid-Level Data Representations

The concept of mid-level data representations provides a convenient way to

characterize certain aspects of signal analysis systems. The analysis process
can be viewed as a sequence of representations from an acoustic signal towards
the analysis result [173], [438]. Usually intermediate abstraction levels are
needed between these two since musical notes, for example, are not readily
visible in the raw acoustic signal. An appropriate mid-level representation
functions as an 'interface' for further analysis and facilitates the design of
efficient algorithms for this purpose.
The most-often used representation in acoustic signal analysis is the short-
time Fourier transform of a signal in successive time frames. Time-frequency
decompositions in general are of fundamental importance in signal processing
and are introduced in Chapter 2. Chapter 3 discusses these in a more general
framework of waveform representations where a music signal is represented
as a linear combination of elementary waveforms from a given dictionary.
Time-frequency plane representations have been used in many transcription
systems (see e.g. [223], [351], [5], and Chapter 9), and especially in percussive

^We may add also that the limitations of human memory and attention affect
the way in which large amounts of data are written down [602].
14 Anssi Klapuri

0.2 0.4 0.6

Time (s]

Fig. 1.6. Three different mid-level representations for a short trumpet sound (FO
260 Hz), followed by a snare drum hit. The left panel shows the time-frequency spec-
trogram with a logarithmic frequency scale. The middle panel shows the sinusoidal
model for the same signal, Une width indicating the amplitude of each sinusoid. The
right panel shows the output of a simple peripheral auditory model for the same
signal.

transcription where both linear [188] and logarithmic [505], [209] frequency
resolution has been used.
Another common choice for a mid-level representation in music transcrip-
tion has been the one based on sinusoid tracks [332], [440], [609], [652]. In
this parametric representation, an acoustic signal is modelled as a sum of
sinusoids with time-varying frequencies and amplitudes [449], [575], as illus-
trated in Fig. 1.6. Pitched musical instruments can be modelled effectively
with relatively few sinusoids and, ideally, the representation supports sound
source separation by classifying the sinusoids to their sources. However, this is
complicated by the fact that frequency components of co-occurring sounds in
music often overlap in time and frequency. Also, reliable extraction of the com-
ponents in real-world complex music signals can be hard. Sinusoidal models
are described in Chapter 3 and applied in Chapters 7 and 10.
In the human auditory system, the signal travelling from the inner ear to
the brain can be viewed as a mid-level representation. A nice thing about this
is that the peripheral parts of hearing are quite well known and computational
models exist which are capable of approximating the signal in the auditory
nerve to a high accuracy. The right panel of Fig. 1.6 illustrates this repre-
sentation. Auditory models have been used for music transcription by several
authors [439], [627], [434], [354] and these are further discussed in Chapter 8.
It is natural to ask if a certain mid-level representation is better than
others in a given task. Ellis and Rosenthal have discussed this question in
the light of several example representations commonly used in acoustic signal
analysis [173]. The authors list several desirable qualities for a mid-level rep-
resentation. Among these are component reduction^ meaning that the number
of objects in the representation is smaller and the meaningfulness of each is
higher compared to the individual samples of the input signal. At the same
 1 Introduction 15

Musicological models,
sound source models

Transcription

Fig. 1.7. The two main sources of information in music transcription: an acoustic
input signal and pre-stored musicological and sound source models.

time, the sound should be decomposed into sufficiently fine-grained elements

so as to support sound source separation by grouping the elements to their
sound sources. Other requirements included invertibility, meaning that the
original sound can be resynthesized from its representation in a perceptually
accurate way, and psychoacoustic plausibility of the representation.

1.2.4 Internal Models

Large-vocabulary speech recognition systems are critically dependent on lan-

guage models which represent linguistic knowledge about speech signals [536],
[316], [321]. The models can be very primitive in nature, for example merely
tabulating the occurrence frequencies of diff'erent three-word sequences (N-
gram models), or more complex, implementing part-of-speech tagging of words
and syntactic inference within sentences.
Musicological information is likely to be equally important for the auto-
matic transcription of polyphonically rich musical material. The probabilities
of different notes occurring concurrently or in sequence can be straightfor-
wardly estimated, since large databases of written music exist in an electronic
format. Also, there are a lot of musical conventions concerning the arrange-
ment of notes for a certain instrument within a given genre. In principle, these
musical constructs can be modelled and learned from data.
In addition to musicological constraints, internal models may contain in-
formation about the physics of musical instruments [193], and heuristic rules,
for example that a human musician has only ten fingers with limited dimen-
sions. These function as a source of information in the transcription process,
along with the input waveform (see Fig. 1.7). Contrary to an individual music
signal, however, these characterize musical tradition at large: its composi-
tional conventions, selection of musical instruments, and so forth. Although
these are generally bound to a certain musical tradition, there are also more
universal constraints that stem from the human perception (see Bregman [49,
Ch. 5]). For example, perceptually coherent melodies usually advance in rel-
atively small pitch transitions and employ a consistent timbre.
Some transcription systems have applied musicological models or sound
source models in the analysis [332], [440], [215], [559]. The principles of doing
16 Anssi Klapuri

this are discussed in more detail in Part IV of this book. The term top-
down processing is often used to characterize systems where models at a high
abstraction level impose constraints on the lower levels [592], [172]. In bottom-
up processing, in turn, information flows from the acoustic signal: features
are extracted, combined into sound sources, and these are further processed
at higher levels. The 'unsupervised-learning' approach mentioned on p. 7 is
characterized by bottom-up processing and a minimal use of pre-stored models
and assumptions. This approach has a certain appeal too, since music signals
are redundant at many levels and, in theory, it might be possible to resolve this
'puzzle' in a completely data-driven manner by analysing a huge collection of
musical pieces in connection and by constructing models automatically from
the data. For further discussion of this approach, see Chapter 9.
Utilizing diverse sources of knowledge in the analysis raises the issue of
integrating the information meaningfully. In automatic speech recognition,
statistical methods have been very successful in this respect: they allow rep-
resenting uncertain knowledge, learning from examples, and combining diverse
types of information.

1.2.5 A Comparison with Speech Recognition

Music transcription is in many ways comparable to automatic speech recog-

nition, although the latter has received greater academic and commercial in-
terest and has been studied longer. Characteristic to both music and speech
is that they are generative in nature: a limited number of discrete elements
are combined to yield larger structures. In speech, phonemes are used to con-
struct words and sentences and, in music, individual sounds are combined to
build up melodies, rhythms, and songs. An important difference between the
two is that speech is essentially monophonic (one speaker), whereas music
is usually polyphonic. On the other hand, speech signals vary more rapidly
and the acoustic features that carry speech information are inherently multi-
dimensional, whereas pitch and timing in music are one-dimensional quanti-
ties.
A central problem in the development of speech recognition systems is the
high dynamic variability of speech sounds in different acoustic and linguistic
contexts - even in the case of a single speaker. To model this variability
adequately, large databases of carefully annotated speech are collected and
used to train statistical models which represent the acoustic characteristics of
phonemes and words.
In music transcription, the principal difficulties stem from combinatorics:
the sounds of different instruments occur in varying combinations and make
up musical pieces. On the other hand, the dynamic variability and complexity
of a single sound event is not as high as that of speech sounds. This has
the consequence that, to some extent, synthetic music signals can be used in
developing and training a music transcriber. Large amounts of training data
can be generated since acoustic measurements for isolated musical sounds are
 1 Introduction 17

available and combinations of these can be generated by mixing. However, it

should be emphasized that this does not remove the need for more realistic
acoustic material, too. The issue of obtaining and annotating such databases
is discussed in [405] and in Chapters 6 and 11. Realistic data are also needed
for the objective evaluation of music analysis systems [465].

1.3 Outline
This section discusses the different subtopics of music transcription and sum-
marizes the contents of each chapter of this book. All the chapters are in-
tended to be self-contained entities, and in principle nothing prevents one
from jumping directly to the beginning of a chapter that is of special interest
to the reader. Whenever some element from the other parts of the book is
needed, an explicit reference is made to the chapter in question.

Part I Foundations

The first part of this book is dedicated to topics that are more or less related
to all areas of music trancription discussed in this book.
Chapter 2 introduces statistical and signal processing techniques that are
applied to music transcription in the subsequent chapters. First, the Fourier
transform and concepts related to time-frequency representations are de-
scribed. This is followed by a discussion of statistical methods, including
random variables, probability density functions, probabilistic models, and el-
ements of estimation theory. Bayesian estimation methods are separately dis-
cussed and numerical computation techniques are described, including Monte
Carlo methods. The last section introduces the reader to pattern recognition
methods and various concepts related to these. Widely used techniques such
as support vector machines and hidden Markov models are included.
Chapter 3 discusses sparse adaptive representations for musical signals.
The issue of data representations was already briefly touched in Section 1.2.3
above. This chapter describes parametric representations (for example the si-
nusoidal model) and 'waveform' representations in which a signal is modelled
as a linear sum of elementary waveforms chosen from a well-defined dictio-
nary. In particular, signal-adaptive algorithms are discussed which aim at
sparse representations, meaning that a small subset of waveforms is chosen
from a large dictionary so that the sound is represented effectively. This is
advantageous from the viewpoint of signal analysis and imposes an implicit
structure to the analysed signal.

Part II Rhythm and Timbre Analysis

The second part of this book describes methods for metre analysis, percussion
transcription, and pitched musical instrument classification.
18 Anssi Klapuri

Chapter 4 discusses beat tracking and musical metre analysis^ which con-
stitute an important subtask of music transcription. As mentioned on p. 10,
metre perception consists of detecting moments of musical stress in an audio
signal, and processing these so that 4;he underlying periodicities are discov-
ered. These two steps can also be discerned in the computational methods.
Measuring the degree of musical emphasis as a function of time is closely re-
lated to onset detection^ that is, to the detection of the beginnings of discrete
sound events in an acoustic signal, a problem which is separately discussed.
For the estimation of the underlying metrical pulses, a number of different
approaches are described, putting particular emphasis on statistical methods.
Chapter 5 discusses unpitched percussion transcription^^ where the aim
is to write down the timbre class, or the sound source, of each constituent
sound along with its timing (see Fig. 1.2 above). The methods discussed in
this chapter represent two main approaches. In one, a percussive track is as-
sumed to be performed using a conventional set of drums, such bass drums,
snares, hi-hats, cymbals, tom-toms, and so forth, and the transcription pro-
ceeds by detecting distinct sound events and by classifying them into these
pre-defined categories. In another approach, no assumptions are made about
the employed instrumental sounds, but these are learned from the input signal
in an unsupervised manner, along with their occurrence times and gains. This
is accomplished by processing a longer portion of the signal in connection and
by trying to find such source signals that the percussive track can be effec-
tively represented as a linear mixture of them. Percussion transcription both
in the presence and absence of pitched instruments is discussed.
Chapter 6 is concerned with the classification of pitched musical instru-
ment sounds. This is useful for music information retrieval purposes, and in
music transcription, it is often desirable to assign individual note events into
'streams' that can be attributed to a certain instrument. The chapter looks at
the acoustics of musical instruments, timbre perception in humans, and basic
concepts related to classification in general. A number of acoustic descriptors,
or features, are described that have been found useful in musical instrument
classification. Then, different classification methods are described and com-
pared, complementing those described in Chapter 2. Classifying individual
musical sounds in polyphonic music usually requires that they are separated
from the mixture signal to some degree. Although this is usually seen as a
separate task from the actual instrument classification, some methods for in-
strument classification in complex music signals are reviewed, too.

Part III Multiple Fundamental Frequency Analysis

The term multiple FO estimation refers to the estimation of the FOs of several
concurrent sounds in an acoustic signal. The third part of this book describes

Many drum instruments can be tuned and their sound evokes a perception of
pitch. Here 'unpitched' means that the instruments are not used to play melodies.
 1 Introduction 19

different ways to do this. Harmonic analysis (writing down the chords of a

piece) can be performed based on the results of the multiple FO analysis,
but this is beyond the scope of this book and an interested reader is referred
to [621, Ch. 6] and [553, Ch. 2]. Harmonic analysis can also be attempted
directly, without note-by-note FO estimation [402], [69], [583].
Chapter 7 discusses multiple FO estimation based on generative models.
Here, the multiple FO estimation problem is expressed in terms of a signal
model, the parameters of which are being estimated. A particular emphasis of
this chapter is statistical methods where the FOs and other relevant parameters
are estimated using the acoustic data and possible prior knowledge about
the parameter distributions. Various algorithms for on-line (causal) and off-
line (non-causal) parameter estimation are described and the computational
aspects of the methods are discussed.
Chapter 8 describes auditory model-based methods for multiple FO esti-
mation. The reader is first introduced with computational models of human
pitch perception. Then, transcription systems are described that use an au-
ditory model as a pre-processing step, and the advantages and disadvantages
of auditorily motivated data representations are discussed. The second part
of the chapter describes multiple FO estimators that are based on an auditory
model but make significant modifications to it in order to perform robust FO
estimation in polyphonic music signals. Two different methods are described
in more detail and evaluated.
Chapter 9 discusses unsupervised learning methods for source separation in
monaural music signals. Here the aim is to separate and learn sound sources
from polyphonic data without sophisticated modelling of the characteristics
of the sources, or detailed modelling of the human auditory perception. In-
stead, the methods utilize general principles, such as statistical independency
between sources, to perform the separation. Various methods are described
that are based on independent component analysis, sparse coding, and non-
negative matrix factorization.

Part IV Entire Systems, Acoustic and Musicological Modelling

The fourth part of the book discusses entire music content analysis systems
and the use of musicological and sound source models in these.
Chapter 10 is concerned with auditory scene analysis (ASA) in music
signals. As already mentioned above, ASA refers to the perception of distinct
sources in polyphonic signals. In music, ASA aims at extracting entities like
notes and chords from an audio signal. The chapter reviews psychophysical
findings regarding the acoustic 'clues' that humans use to assign spectral
components to their respective sources, and the role of internal models and
top-down processing in this. Various computational approaches to ASA are
described, with a special emphasis on statistical methods and inference in
Bayesian networks.
20 Anssi Klapuri

Chapter 11 discusses a research approach called music scene description^

where the aim is to obtain descriptions that are intuitively meaningful to an
untrained listener, without trying to extract every musical note from musical
audio. Concretely, this includes the analysis of the melody, bass lines, met-
rical structure, rhythm, and chorus and phrase repetition. In particular, two
research problems are discussed in more detail. Predominant FO estimation
refers to the estimation of the FO of only the most prominent sound in a poly-
phonic mixture. This closely resembles the experience of an average listener
who catches the melody or the 'theme' of a piece of music even though he
or she would not be able to distinguish the inner lines. Here, methods for
extracting the melody and the bass line in music recordings are introduced.
The other problem addressed is music structure analysis, especially locating
the chorus section in popular music.
Chapter 12 addresses singing transcription, which means converting a
recorded singing performance into a sequence of discrete note pitch labels
and their starting and ending points in time. The process can be broken into
two stages, where first a continuous track of pitch estimates (and possibly
other acoustic features) is extracted from an acoustic signal, and these are
then converted into a symbolic musical notation. The latter stage involves the
segmentation of the pitch track into discrete note events and quantizing their
pitch valuestasks which are particularly difficult for singing signals. The
chapter reviews state-of-the-art singing transcription methods and discusses
the use of acoustic and musicological models to tackle the problem.
An Introduction to Statistical Signal
Processing and Spectrum Estimation

Manuel Davy

LAGIS/CNRS, BP 48, Cite Scientifique, 59651 Villeneuve d'Ascq Cedex, Prance

Manuel.DavyOec-lille.fr

This chapter presents an overview of current signal processing techniques,

most of which are applied to music transcription in the following chapters.
The elements provided will hopefully help the reader. Some signal processing
tools presented here are well known, and readers already familiar with these
concepts may wish to skip ahead. As we only present an overview of various
methods, readers interested in more depth may refer to the bibliographical
references provided throughout the chapter.
This chapter is organized as follows. Section 2.1 presents the Fourier trans-
form and some related tools: time-frequency representations and cepstral coef-
ficients. Section 2.2 introduces basic statistical tools such as random variables,
probability density functions, and likelihood functions. It also introduces es-
timation theory. Section 2.3 is about Bayesian estimation methods, including
Monte Carlo techniques for numerical computations. Finally, Section 2.4 in-
troduces pattern recognition methods, including support vector machines and
hidden Markov models.

2.1 Frequency, Time-Frequency, and Cepstral

Representations
As explained in Chapter 1, the two main domains where musical signals can
be represented and understood are time and frequency. If time is the domain
where the music signals are played and recorded (air pressure vs. time data),
frequency is the domain where they can be represented and understood: in
Western music, the height of a note in a score represents its fundamental fre-
quency. From a mathematics viewpoint, frequency is defined via the Fourier
transform (FT), which is introduced in Section 2.1.1. However, the Fourier de-
finition of frequency is not practical: frequency is well defined only for infinite
length, stationary, continuous, pure sine wavesthese objects do not belong
to the real world, and non-stationarity is everywhere. The FT nevertheless
provides a lot of useful information.
22 Manuel Davy

If time alone or frequency alone are not enough to represent music, then
we need to think in terms of joint time and frequency representations (TFRs).
Western musical scores are actually TFRs with a specific encoding. Such rep-
resentations are introduced in Section 2.1.2, and rely on the concept of/mme,
that is, a well time-localized part of the signal. From these frames, we can de-
fine a simple TFR, the spectrogram, as well as a time and time-lag represen-
tation called cepstral representation; see Section 2.1.3. The following section
discusses some basic properties of Fourier Transforms.

2.1.1 T h e Fourier Transform

The frequency representation of signals is given by the Fourier transform^

(FT). In practical situations we generally have to deal with sampled, real-
valued discrete time signals, denoted x(n), where n denotes discrete time.
The corresponding discrete FT is defined in (2.1) where k denotes the dis-
crete frequency. It is sometimes more convenient to define music transcription
methods in terms of continuous time signals, denoted x{t), where t denotes
continuous time. The continuous FT is defined in (2.2), and / denotes con-
tinuous frequency.
oo

DFT^(A;) = X{k) = ^ x(n)e-J'2'^'=", (2.1)

n= oo
oo
x{t)e-^^^f'dt. (2.2)
/ -OO

The continuous and discrete FTs map the signal from the time domain to the
frequency domain; X[f) and X{k) are generally complex valued. The inverse
Fourier transforms (IFTs) are also quite useful for music processing; they are
defined in (2.3) and (2.4) below.
oo

IDFTx(n) = ^ X(fc)e^'2''=" = x{n), (2.3)

fc= O O
OO

X{f)e^^^f'df = x{t). (2.4)

/ -OO

An efficient approach to computing DFTs (2.1) and IDFTs (2.3) is the fast
Fourier transform (FFT) algorithm; see [531].
Some properties of the F T / I F T are of importance in music transcription
applications. In particular, the FT is a linear operation. Moreover, it maps
the convolution operation into a simple product.^ In other words, considering
^Readers interested in more advanced topics may refer to a dedicated book;
see [48] for example.
^The convolution operation is used for signal filtering: applying a filter with
time impulse response h{n) to a signal x{n) is done by convolving x with h or,
equivalently, by multiplying their FTs.
 2 An Introduction to Signal Processing 23

for example the discrete time case, DFTa;^*x2(^) = DFTxi(fc) x DFTa;2(^),

where * denotes the convolution defined as follows:
oo

xi(n) * X2{n) = 2_] xi{i)x2{n i) (discrete time signals), (2.5)

i = cx)
oo
xi{r)x2{t T)dr (continuous time signals). (2.6)
/ -CX)

This result is also true for the continuous-time FT, as well as for the inverse
transforms. Another important property, the Shannon theorem, states that
the range of frequencies where the discrete F T is meaningful has an upper
limit given by kg/2 (this is the Nyquist frequency, where /cg is the sampling
frequency). A straightforward consequence of the Shannon theorem is that
digital signals sampled at the CD rate k^ = 44100 Hz can be analysed up to
the maximum frequency 22050 Hz.

2.1.2 Time-Frequency Representations

There exists only one Fourier transform of a given signal, as defined above.
However, there is an infinite number of time-frequency representations (TFRs).
The most popular one is the spectrogram, defined as the Fourier transform of
successive signal frames.^
Frames are widely used in audio processing algorithms. They are portions
of the signal with given time localizations. More precisely, the frame localized
at time to, computed with window w and denoted 5j^(r), is

slit) = x{tMto-t), (2.7)

where w is a window; see Fig. 2.1. Windows are generally positive and symmet-
ric (that is, w(t) w(f), for all t) with a limited support (i.e., there exists
a time ti such that w{t) = 0, for all t > t i ) . Standard window shapes are
Gaussian, Hamming, Hanning, or rectangular^ and typical frame durations
are from 20 ms to 100 ms in audio processing. As a rule of thumb, rectangular
windows should not be used in practice, except under special circumstances.
From frames, it is easy to build short time Fourier transforms (STFTs) as the
FT of successive frames:
oo
(2.8)
/ -OO

^We restrict this discussion to continuous time signals, for the sake of simplicity.
Discrete time-frequency representations of discrete signals are, in general, more com-
plex to write in closed form, and they are obtained by discretizing the continuous
representations; see [191].
^A list of windows and reasons for choosing for their shapes can be found in [244],
[61].
24 Manuel Davy

Signal x(t)

Window w(^Q - 1 )

Frame s^ (t)

time t
Fig. 2.1. The frame SYQ is obtained by multiplying the signal x by a sliding window
w centred at time t.

Spectrograms are energy representations and they are defined as the squared
modulus of the STFT:

SP:(t,/) = |STFT:(f,/)|^ (2.9)

Changing the window w defines a different STFT and thus a different spec-
trogram. Any spectrogram can also be interpreted as the output of a bank
of filters: considering a given frequency / , S P ^ ( t , / ) is the instantaneous en-
ergy at time t of the output of the filter with frequency response given by
CFTw(/ i^), i^ ^ [00,00], applied to the signal x. As a consequence, a short
duration window leads to a spectrogram with good time resolution and bad
frequency resolution, whereas a longer window leads to the opposite situa-
tion.^ Figure 2.2 represents two spectrograms of a piano excerpt, illustrating
the influence of the window length on the representation.
Another interpretation of the STFT arises when considering (2.8) as the
dot product between x(r) and the windowed complex sinusoid w{tr)e~^'^^^'^:
under some conditions, the family of elementary time-frequency atoms {w(^
^jg-j27r/T j ^ ^ forms a basis called a Gabor basis. In such cases, STFTJJ" is a
decomposition of x on this basis, yielding a Gabor representation of x; see
Chapter 3 and [191], [184].
Spectrograms being energy representations, they are quadratic in the sig-
nal X. They also have the time-frequency covariance property: Let us shift x in
time and frequency, defining xi{t) x{t to) exp{j27r/o}. The time-frequency
covariance property ensures that SP^^ (t, / ) = SP^ (t - to, / - /o) Many other

^A TFR is said to have good time (frequency) resolution if the signal energy is
displayed around its true location with small spread along the time (frequency) axis.
The product of time resolution and frequency resolution is lower bounded via the
Heisenberg-Gabor inequality [191].
 2 An Introduction to Signal Processing 25

Time (s^ ^-^ Time (s)

Fig. 2.2. Spectrograms of a piano excerpt computed with Hamming windows with
length 20ms (left) and 80ms (right). The time-frequency resolution is highly depen-
dent on the window length.

quadratic and covariant time-frequency representations can be defined. They

are all derived from the Wigner-Ville representation:

WV,(f,/) = P ^{t+D ^* {t - 0 e-^^'f'dr, (2.10)

where x^ is the complex conjugate of x, via a time and frequency convolution

with a time-frequency kernel (/)(t, / ) [191]. In particular, the spectrogram is ob-
tained by using ^{t, f) = WVw(^, / ) ; that is, SP:(t, / ) = [WV, * WVw] (t, / ) ,
where * denotes the time-frequency convolution and WVw is the Wigner-Ville
representation of the window.

2.1.3 The Cepstrum

Filtering a signal is actually a convolution operation in the time domain, and

a product operation in the frequency domain. The aim of cepstral represen-
tations was initially to turn filtering into an addition operation by using the
logarithm of the FT. More precisely, the cepstrum of a continuous signal x(t),
with FT denoted X ( / ) , is defined as

Cep,(r) = ICFT log(|X|) (r) log(|X(/)|)e^^''^^rf/. (2.11)

J C

and we see that filtering a signal in the frequency domain through the product
X{f)H{f) becomes, after taking the log, the addition log[X(/)] -f log[if(/)]
where H{f) is the filter frequency response. When dealing with discrete time
signals, the cepstrum is also discrete and is generally referred to in terms of
cepstral coefficients. Cepstral coefficients can be computed from the discrete
26 Manuel Davy

50 ^^w.^^<|# <'iii..i,i'' ' ' t w O ^ i i

30 M**^!*****' 'twx^*'**^"**"*" ^ * " '

20 M A ^iMWN^iiMlwiKu^O'illiMiiiiKiii.^ i[^.><IW<"** '

10 Lr*,!,^^,^^

fl
^
S 2

0.5 1.5
Time (s)

Fig. 2 . 3 . Time evolution of the first cepstral coefficients for the piano signal of
Fig. 2.2. The coefficient number is written along the vertical axis. For the sake of
figure clarity, the magnitude of the first cepstral coefficient has been divided by five,
and an offset has been added to each coefficient to avoid overlap.

equivalent of (2.11). Figure 2.3 displays t h e time evolution of some of t h e first

cepstral coeflficients for a music excerpt.
In practice, one uses linear prediction cepstral coefficients (LPCCs) or mel
frequency cepstral coefficients (MFCCs) [536]. L P C C s are computed by ap-
plying a linear prediction model (also called an autoregressive model; see Sec-
tion 2.2.9, p. 39) to the signal and transforming the resulting coeflBcients into
L P C C s . M F C C s are used much more in music processing, and the steps of
their computation are plotted in Fig. 2.5. T h e input signal x{n) is first win-
dowed, resulting in frames 5^^ (n) centred at time no- T h e magnitude spectrum
of each frame, denoted \S^^{k)\, is sent into a bank of K^^i mel frequency fil-
ters (typically K^^i = 40). Mel frequency filters have a triangular shape in
t h e frequency domain; see Fig. 2.4. T h e central frequencies of the filters are
equally spaced in terms of mel frequencies, which relate to the standard, linear
frequency t h r o u g h t h e relation [297]

/c^ei = 2595 log 10 +1 (2.12)

700

For each mel filter, frequency components within its passband are weighted
by the magnitude response of the filter, and then squared and summed. The
resulting filter-related coefl&cient is denoted Xno(^mei)- For the full set of fil-
ters, the Xno(^mei)'s are stacked into a vector of size K^^i, whose logarithm
 2 An Introduction to Signal Processing 27

Frequency
Fig. 2.4. Mel filter bank. Each filter has a triangular shape, and unity response
at its centre. Its edges coincide with the adjacent filters' central frequencies. The
central frequencies are linearly spaced on the mel frequency scale, which results in an
exponential interval between the filter centres onto the linear scale, through (2.12).

is transformed back into the time lag domain using the discrete cosine trans-
form ( D C T ) , where the D C T of a discrete signal x with length T is defined
as follows:
T
DCT^(i) = ^x{n)cos\-iin--j (2.13)

In addition to yielding time-domain coefficients, the D C T also decomposes

the coefficients in a way similar to principal component analysis [417] (see
p. 54), so t h a t each mel cepstral coefficient carries diflFerent information from
the other M F C C s . In many applications, one is also interested in the time
evolution of M F C C s , which can be given as derivatives. T h e first and second
derivatives for each M F C C Cep^^{i) {i = l , . . . , K ^ e i ) at each time no are
denoted A n o ( 0 and AAno{i).

x{n) Sno(n) \SZ{k)\

sliding mel filter
^
window w(n no)
IDFTI
bank
El
(/Cmel)

10g(Xno(^mel))
DCT logarithm
CepJi)^

differentiator

differentiator

Fig. 2.5. MFCC computation steps. The magnitude spectrum |5no(^)| of each
frame S^Q (n) is filtered through the mel filter bank. The squared outputs of each
filter are summed over the filter frequency range, yielding a coefficient Xno(^mei) for
each filter. The vector made of the logarithm of these coefficients is mapped back
to the time domain using the discrete cosine transform.
28 Manuel Davy

2.2 Basic Statistical Methods

Musical data have some common characteristics, but they may also vary a
lot from one instrument to another, or because of changing recording condi-
tions. Statistical methods are very powerful for modelling sound sources and
other entities in music because they can handle their variability. Even a single
sound source produces a great variety of sounds, and it cannot be modelled
deterministically. The main concept in statistical methods is that of random
variables which are characterized by a probability density function (pdf).

2.2.1 Probability Distributions and Density Functions

Roughly speaking,^ a random variable a is a scalar or a vector in some space

A which can be continuous (e.g., ^ = R) or discrete (e.g., A = Z). It is
called random because it can take several values in A which are impossible to
predict without any errors. If random variables cannot be predicted, they can
be characterized, since some values of a may be more likely to appear than
others. In the case where a is a discrete variable, it is characterized by the
probabilities that each value of a will appear, denoted P(a). For example, in
the coin tossing problem, A = {tail, head} and the probabilities are P(a =
head) = P(a = tail) = 1/2. Standard discrete random variables distributions
are uniform (like in coin tossing), Poisson, binomial, etc. [544].
In the continuous case, it may happen that all values of a are equally hkely
in a part B of the space A, and never appear out of B. In that case, a is a
continuous random variable with a uniform pdf denoted UB, with

^^B(a) = ^ ] l B ( a ) , (2.14)

where /i(S) is called Lebesgue measure of B (it measures the volume of B) and
Is is the indicator function which satisfies I^(a) = 1 if a G B and I^(a) = 0
otherwise. Uniform pdfs are often met in audio processing problems, and they
are used to model the lack of precise information about a parameter.
Another important pdf in engineering problems is the Gaussian^ also called
normal pdf.

M{a;fi,S) = [det(27ri:)]-i/'exp|-i(a-/x)Ti:-i(a-^)|, (2. 15)

where /i is the mean vector^ and U is the covariance matrix, which is sym-
metric. The vector /i has the same size as a, and the matrix U is square, its
size in both dimensions being that of a. An exact definition of /x and U is
given in (2.19) and (2.20). An important special case is when a is scalar. In
this case, the Gaussian pdf is written as
The precise definition of random variables is beyond the scope of this chapter;
see [548] for a more precise introduction.
 2 An Introduction to Signal Processing 29

J\f(a;m,a^) = exp (2.16)

\/27rcr

where m is the mean and cr^ is the variance. There exist many other standard
continuous pdf shapes such as the gamma, inverse gamma, Cauchy, Laplace,
Dirichlet, etc. [544]. When the random variable a is a vector with dimension
da, a = [ a i , . . . , adj^, then the pdf p(a) is in fact a joint pdf p ( a i , . . . , ad^)-
This is a function of dimension da and it gives us the probability of observing
jointly ai, . . . , a^^.
Two very important properties of pdfs and discrete variable distributions
are that they are always positive and their sum equals one:

^ P(a) = 1 (discrete case), (2.17)

L p(a) da.= 1 (continuous case). (2.18)

2.2.2 Mean, Covariance, Expectation, Moments, and Modes

For any continuous pdf p(a), it is generally possible to define the mean vector
/i and covariance matrix U as follows:

/x = / a p(a) da, (2.19)

JA

E= / (a - /x)(a - M)"^ p(a) da. (2.20)

JA
Note that both /x and U cannot be defined whenever the integrals in (2.19)
and (2.20) do not converge.
For both discrete and continuous random variables, the mean and covari-
ance matrix can be expressed in terms of the expectation E of a function h
over A, defined as follows:

Ep(a)[/i(a)] = ^ h{a)P{3i) (discrete case), (2.21)

aeA

Ep(a)[/^(a)] = / /i(a)p(a)da (continuous case). (2.22)

JA
The expectation can be interpreted as follows: consider for example the con-
tinuous case, and assume that a set of random variables { a i , . . . , a i v } are
identically distributed according to some pdf p(a); then the so-called empiri-
cal average
1 ^
Ai^(Mai),...,Maiv)) = ^E'^^^^) (2.23)
2=1
30 Manuel Davy

is an approximation of Ep(a)[/i(a)] in the sense t h a t Jif^{h{aii)^... ,h{3.N))

converges to Ep(a)[ft'(a)] when the number of random variables A^ becomes
infinite (this is t h e law of large numbers). It is easy t o understand t h e expec-
tation as the average of many samples h{a.) where a is distributed according
to p(a) (this is also true for discrete r a n d o m variables).
Using the expectation operator for any continuous or discrete r a n d o m
variable, the mean vector and covariance matrix can be written as follows:

/x = Ep(a)[a] and 27 = Ep(a)[(a - / x ) ( a -/x)"""] (discrete case), (2.24)

/x = Ep(a) [a] and 27 = Ep(a) [(a -/LA)(a -/x)"*"] (continuous case). (2.25)

More generally, the order r moment Mr oi di scalar r a n d o m variable a with

mean m is given by the expectation

Mr [P{a)] = Ep(^) [(a - my] (discrete case), (2.26)

Mr [p(a)] = Ep(a) [{a - my] (continuous case). (2.27)

In t h e case where the random variable a is a vector of dimension c?a, its order
r moment A l ^ is an r-dimensional tensor with size c^a x da x . . . x da-
Finally, distributions generally have modes^ t h a t is, local maxima: their lo-
cations indicate in which regions of the space A the r a n d o m variable a is more
likely to appear. Figure 2.6 summarizes graphically t h e concepts introduced
in this subsection.

0.4

3.

a(2) .,4 : 4 1
ch-1
a(l)
One-dimensional Gaussian Two-dimensional Gaussian

Fig. 2.6. Graphical representation of the Gaussian pdf of a scalar (continuous)

random variable (left) and a two-dimensional Gaussian random variable (right). In
the scalar case (left), the mean and standard deviation are equal to one. In the two-
dimensional case (right), the mean vector is [-1 1.2]^ and the covariance matrix
is 27(1,1) = 1, 27(1,2) = 27(2,1) = 0.6 and 27(2,2) = 1.2. The diagonal terms
of the covariance matrix being non-zero, the pdf is tilted around the mean point
(-1,1.2). At the back of the right figure, the marginal pdfs (see Section 2.2.3) have
been plotted.
 2 An Introduction to Signal Processing 31

2.2.3 Marginal and Conditional pdfs

Considering two random variables ai G Ai and a2 G .42, we can define their

joint pdf p(ai, a2), which informs us about the probability of observing jointly
(ai,a2). Then we can define the marginal pdfs,

p(ai) = / p(ai,a2)da2 and p(a2) = / p(ai,a2) c^ai, (2.28)

JA2 JAI

which can be interpreted as the pdf of one of the random variable irrespective
of the value of the other one; see Fig. 2.6, right. Using marginal pdfs, it is
possible to decompose the joint pdf p(ai,a2) as follows:

p(ai,a2) = p(ai|a2)p(a2) = p(a2|ai)p(ai), (2.29)

where p(ai|a2) is the conditional pdf of a i , and should be read as 'the pdf of
ai conditional on a2'. Its interpretation is simple: Imagine that a2 has some
fixed, non-random value denoted a2'^''. Then ai is still a random variable,
and its pdf is p(ai,a2 = a2'^'') up to a normalizing constant. This constant is
necessary because the integral of p(ai,a2 = ^2'^'^) with respect to (w.r.t.) ai
does not equal one anymore. The notation p(ai|a2) should be understood as
p(ai|a2'^'^), that is, (1/C)p(ai,a2 = a2'^'^) where the normalizing constant C is
p(a^^^^) from (2.29).
Finally, two random variables ai and a2 are independent if and only if
p(ai,a2) = p(ai) p(a2) or, equivalently, p(ai|a2) = p(ai) or p(a2|ai) = p(a2).
This means that the knowledge of ai provides no information about a2, and
vice versa.

2.2.4 Probabilistic Models

Random variables are extremely useful in signal processing because they can
model the lack of certainty about a physical phenomenon. Imagine we have a
good model for some process: for example, a 'pure sine' acoustic waveform gen-
erated by an electronic instrument. A model for the pressure signal recorded
is given by the following discrete time sine model:

x{n) = asin(27rA:on-f-(/)o) for n = l , . . . , T , (2.30)

where /CQ is the unknown sine waveform frequency, 0o is the initial phase, and
a is the signal amplitude. It is clear that the recorded pressure signal will not
fit exactly the model (2.30), and that it will deviate from it. As these deviations
may have various causes (air temperature/pressure inhomogeneity, non purely
sinusoidal loudspeaker behaviour when emitting the sound, digital to analog
conversion artifacts, etc.), it is unrealistic to model them deterministically,
and a random model can be used. A possible such model is

x(n) = asin(27rfcon + (/>o) + e(n) for n = l , . . . , T , (2.31)

32 Manuel Davy

where e(n) is a so-called additive random noise with a given pdf. Including
this noise in the model is aimed at modelling the deviations of the recorded
data from the model in (2.30). In the general case, it is assumed that e(n) is
a stationary white noise, or independent identically distributed (i.i.d.) noise-,
that is, the noise at any time ni is statistically independent of the noise at
any time n2, and their pdfs are equal. More precisely, the joint pdf of the
noise samples equals the product of the pdfs of each sample: Writing the
noise samples as a vector e = [e(l),..., e(T)]^ for n = 1 , . . . , T, we have
p(e) = p{eil))p{e{2)).. . p ( e ( r ) ) , with p(e(l)) - p{e{2)) =...= p{e{T)).
Finally, since the noise is aimed at modelling small deviations around the sine
model, it is assumed to be zero-mean. Equation (2.31) defines a probabilistic
signal model and directly yields the likelihood function.

2.2.5 Likelihood Functions

In (2.31), it is assumed that the recorded signal x follows a sine model with
additive zero-mean white noise. In practice, this model is interesting in the
sense that it relates the recorded signal x(n), n 1 , . . . , T to the parameters
fco, 00, and a. In the following, we denote by 6 the set of unknown parameters,
i.e., 6 = [ko,(l)o,a]. From the probabilistic model defined above, and given a
recorded signal x, we see that some values of 0 are more likely than others:
for example, if the sine wave is generated with frequency 440 Hz, finding ko =
440 Hz is very likely. It is also likely that the loudspeaker which emits the
sound adds partials (that is, additional sine waves with lower amplitudes)
at frequencies 880 Hz, 1320 Hz, etc. These are also hkely, to a lower extent
than 440Hz, though. Conversely, assume 0 is given; then the signal x(n),
n = 1 , . . . , T can be seen as a random vector denoted by x = [x(l), ,x{T)]^.
It admits a joint pdf p(x|0) = p(x(l), x ( 2 ) , . . . , x{T)\6), conditional on 6: by
changing 0, the signal pdf p(x|^) is changed. In the sine example presented
above, assuming the noise is Gaussian, the covariance matrix of e is diagonal
of size T, with the variance a"^ of each e(n), n = 1 , . . . , T on its diagonal. This
pdf is

p(x|0) = [ 2 7 r a 2 ] - ^ / 2 e x p | - ^ [ x - f W ] ' ' [ x - f ( 0 ) ] | ^232)

where f{6) is the model given in (2.30) written in vector form, i.e., f(0)
[asin(27rA:ol + (/>o), ., a sin(27r/coT -j- ^o)]""".
The mathematical object p(x|^) admits two interpretations: First, when
read as a pdf of x for a given 6, p(x|0) is called the conditional pdf of x^
conditioned on 6. Second, when seen as a function of 6 for given x, p(x|0)
is called the parameter likelihood function and it is defined over the space of
all possible values of 9 denoted O. For example O = [0,1/2] x [0,27r] x [0, oc)
in the sine example (2.30). Note that the function p(x|^) is not a pdf of 6,
because its integral w.r.t. 0 over 0 may not equal one.
 2 An Introduction to Signal Processing 33

In practice, we always refer to p(x|0) as the likelihood function. It should

be interpreted as a measure of similarity between the model f (0) tuned by
0 and the signal x: the larger the likelihood, the more likely the parameter
vector 6. In some problems, it can happen that the likelihood function takes
infinite values for some values of 6. In this case, it is called degenerate.

2.2.6 Maximum Likelihood Estimation

As outlined above, the Hkelihood can be seen as a similarity measure be-

tween the signal and the model, via its parameters. In the sine example given
in (2.30), the parameter vector 0 = [fco, 00, <^] is unknown and it may be useful
to estimate it. It is quite natural to use as an estimate the value of 6 that
maximizes the likelihood (assumed non-degenerate):

0ML = argmax p(x|0), (2.33)

eee
where the subscript ML stands for maximum likelihood (ML). Figure 2.7 shows
the likelihood function corresponding to the model (2.31), and the maximum
can be seen: it is located at point ^ML- The optimization problem (2.33) is often
transformed into the equivalent problem of maximizing the log-likelihood^
defined as follows:
Ue) = log[p(x|0)]. (2.34)
Thanks to their excellent properties (see subsection below), ML estimators
are often used in signal processing applications.

2.2.7 Elements of Estimation Theory

There exist many estimation techniques with possible apphcation to music

transcription: Aside from maximum likelihood, one can quote minimum mean
square error (MMSE) estimators;^ they consist of estimating the expecta-
tion Ep(^) [h{0)] of some function h of the parameter by using an empiri-
cal average (2.23). Estimators are characterized by two important quantities:
the bias and the estimator covariance. In the sine wave example presented
above, the sound produced by the electronic instrument through a loudspeaker
can be recorded several times, resulting in several discrete signal vectors x^,
i = 1 , . . . , / , all described by the model (2.31) (this is because the signal is
random). When applying some estimation technique on each of these signals,
such as ML estimation, we obtain estimates 0^, i = 1 , . . . , / , which all have a
different value. In other words, the estimate 0 is a random variable with pdf

^The logarithm function is order preserving, and thus maximizing a function or

maximizing its logarithm yield the same result.
^MMSE estimates are further considered in Section 2.3.4, p. 41.
34 Manuel Davy

From above 3D view

Fig. 2.7. Likelihood p(x|fco, </)o, Q;) in (2.32) as a function of the frequency ko and
the amplitude a for a sine wave with frequency ko = 0.3 and amplitude a = 2. The
sine wave is corrupted with a white Gaussian zero-mean additive noise with variance
a^ = 4. Both this variance and the initial phase (^o are assumed to be known. The
likelihood shows a very sharp peak, which leaves no doubt about the parameter
value, but which may be hard to localize.

P u t t i n g aside estimation issues, and assuming the model (2.31) is correct,

the recorded signals x^ are all generated by one value of 0, denoted 0gt,
where the subscript 'gt' stands for 'ground t r u t h ' : this is the value we want
to estimate.
In t h e ^ e n e r a l case, we define the bias of the estimator which produces the
estimate 0 of 0 as follows:

Bias() = E p ( ^ ) ( ^ ) - ^ g t . (2.35)

Note t h a t the bias characterizes the estimator (e.g., the ML estimator applied
to the model (2.31) in the sine wave example) and not an estimate 0. An
estimator is said to be unbiased whenever Bias(0) = 0, and this is, of course,
a very important property.
T h e estimator covariance matrix Var(^) is defined for b o t h biased and
unbiased estimators (in the latter case, it is also referred to as the mean
square error):
Var() - E[(e -E[d])(e -E[d]y]. (2.36)

For unbiased estimators, Var(^) is lower bounded by the so-called C r a m e r -

Rao bound.^ W h e n the information about the signal becomes stronger,

^The definition of a lower bound for scalar numbers has to be adapted to matrices:
A matrix A is a lower bound for the covariance matrix U if and only if the matrix
E A is a, definite non-negative matrix (where A and U have the same size);
see [295].
 2 An Introduction to Signal Processing 35

typically because we have longer records (T -^ CXD), we can study the estima-
tor asymptotic properties. In particular, the estimator is said to be consistent
whenever the variance Var(0) tends to zero.
ML estimators are asymptotically unbiased (in some cases, they are even
unbiased for any number T of data); they are consistent; and they reach
asymptotically the Cramer-Rao bound. Because of these excellent properties,
maximum likelihood approaches are widely used in signal processing. When
a good model is defined, the implementation difficulty consists mainly of the
optimization problem (2.33), which can usually be solved by the expectation-
maximization algorithm.

2.2.8 The Expectation-Maximization Algorithm

The expectation-maximization (EM) algorithm is an optimization algorithm

specially designed for ML parameter estimation. Before describing it, we em-
phasize two of its important characteristics: First, it is 'gradient-based'; that
is, it does not find the global maximum in the general case, unless conve-
niently initialized. Second, it is only applicable to problems with latent vari-
ables. There are a great number of problems which can be addressed using
EM. As an illustration, we consider here the important problem of fitting a
mixture of Gaussians to some data.^^

A n Example: Gaussian Mixture Models (GMMs)

Gaussian mixture models are quite important in audio processing. For ex-
ample in speech processing the data considered are generally sets of cepstral
coefficients. However, this algorithm is not restricted to cepstral data. In gen-
eral, GMMs are used to estimate the pdf of a set of data. This is because they
have two key properties: 1) given that there are enough Gaussians in (2.37),
GMMs can approximate any pdf (versatility), and 2) finding their parameters
is easy thanks to the EM algorithm.
Consider the set of data X = { x i , . . . , Xm}, where each individual datum
is a vector in R^^. In the speech processing example mentioned above, each
datum Xi {i = l , . . . , m ) is made of the first dx cepstral coefficients of an
audio signal frame. Gaussian mixture modelling consists of fitting a mixture
pdf made of J Gaussians on each datum in X. The mixture pdf is
J
p(x,|{/?M.,i;,},-=i,...,j) - ^/?,Ar(x,;Ax,-,^,), (2.37)

where the mixture coefficients /?j, j = 1 , . . . , J are such that Ylj=i 0j = 1.

The likelihood of the mixture for the complete set X is written as
^^Further details about the EM algorithm for general purposes may be found
in [471] and the special case of Gaussian mixtures is addressed in [536], [38].
36 Manuel Davy

m
(2.38)
i=l

because the d a t a in X are assumed to be independent and identically dis-

tributed (thus, the joint dataset pdf is the product of the individual d a t u m
pdfs).

1 1
1 1
1 1
1 1
^ 1
1
1
1
^ 1 1
I 1
I I /'
1 1 /
' ; /
/ \ ; '
/ X 1 '

r i - - -^.
...^ .. *_
-10 -6 -3 -0.2 2 10
a
Fig. 2.8. Mixture of five Gaussians. The five Gaussians composing the mixture
are represented in dotted Hues, whereas the full mixture pdf is represented in solid
Hne (with offset -hl.l for better visibility). The mixture coefficients are pi = ... =
05 = 1/5. The means of the Gaussians are [6, 3, 0.2, 2, 5] and the variances are
[1, 0.16, 0.64,1, 3.24]. Any random variable distributed according to this mixture pdf
may be generated either directly from the mixture pdf, or by first selecting randomly
one of the five Gaussians, and then generating the variable from it. For example,
the dot at abscissa 1.8 can either result from direct sampling form the mixture in
solid fine, or by first selecting randomly one of the Gaussians (here, the Gaussian
centred which mean equals 2), and samphng from this Gaussian.

In Gaussian mixture approaches, X is described statistically t h r o u g h the

mixture parameters {/?j,/x^, 27j}j=i,...^j. This modelling approach is also in-
teresting because it compresses the information of m d a t a into J parameter
sets {Pj, fip 17j}, with J <^m. These parameters need to be estimated, how-
ever. This is implemented here by the E M algorithm for ML estimation of the
unknown parameters, denoted for convenience by 0 = { ^ i , - - - , ^ j } , where

In order to apply the E M algorithm, we note t h a t the mixture (2.37)

admits two interpretations: First, the pdf of x^ is seen as a sum of weighted
Gaussians, as depicted in Fig. 2.8. Second, with probability (3j (j = 1 , . . . , J ) ,
the d a t u m x^ is distributed according to the pdf A/'(xi; /x^, ^j)- In the second
 2 An Introduction to Signal Processing 37

interpretation, we can associate the latent variables z^, z = 1 , . . . , m to each

datum Xi, where zi takes its values in { 1 , 2 , . . . , J } . The latent variable zi
indicates the component N{^i\ /x^, Ej) of the mixture (2.37) that corresponds
to x^. In other words, when zi is known, the distribution of Xf is
p{^i\zi^j,Pj,^^,Ej), - A/'(xi;/i^.,^,). (2.39)

The latent variable Zi probability is P(2:^ = j ) = /3j.

Principle of the EM Algorithm

The general structure of the EM algorithm is presented in Algorithm 2.1
below. It is assumed that the model involves the data set X and latent variables
Z= {zi,...,zj}.

Algorithm 2.1: General EM Algorithm

1. Initialization: Select the starting point 0^^^ inside O.
2. Iterations: For j = 1, 2 , . . . , do (until convergence is reached)
Set
^(^+1) = argmaxQx(^|6/^^^), (2.40)
e
where
Qx{e\0^'^) = Ep(x,e(.))[logp(X,Z|^)]. (2.41)

Algorithm 2.1 is quite simple. The two steps used to compute 0^-^'^^^ from
e^^^ are
The expectation step: compute Qx(^|^ ) using (2.41).
The maximization step: maximize (5x(^|^ ) with respect to 0. This yields
00+1).
The principle of this algorithm [471] is that we want to maximize the log-
likelihood log p(X, Z\6) with respect to 6 without knowing the latent variables
Z. The expectation in (2.41) permits us to get rid of the latent variables, so
as to perform the maximization. However, since this expectation is computed
for the parameter value 6^-^^, which is not the 'true' value, the expectation in
(2.41) is not the 'true' one. The iterations in Algorithm 2.1 ensure that 0^^^
becomes closer to the 'true' value, yielding a better approximation of the true
expectation in (2.41), and thus a better ML estimate at each iteration.

Application to the Gaussian Mixture Model

In the Gaussian mixture case, (2.41) and (2.40) can be calculated analytically.
m
log piX,Z\e) = ^ l o g pixi\zi,d) + \og P{zi\e), (2.42)
2=1
38 Manuel Davy

and, using (2.41) (see [38] for details),

J m
Qx{0\e^'^) - 5 ] 5 ^ log [^,p(x,|z, = j , 0 , ) ] P ( z , - j | x , , 0 f ) , (2.43)

where P{zi = j\:s.i,6j) is computed using Bayes's rule; see (2.51) in Sec-
tion 2.2.3, p. 31. Moreover, Qx{0\e^^^) in (2.43) can be maximized analyt-
ically. Defining S. ^u) = J2^i ^i^i Jl^i^^j )^ ^^^ ^ ^ update equations
are [38]

^(i+i) 1
^j=7;;^.ov^^ (2-44)
m

^ r ' ' = ^"o) J2'^Pizi-j\^i,ef), (2.45)

m

^r'^ = S'la^ E [^ - ^r''] [^^ - t^^'^V^i^i = il^^. ^?)- (2-46)

' i = l

An interesting remark is that the likelihood of Gaussian mixture models is

degenerate. In particular, when there are more components than data points
(i.e., when J > m)^ the ML estimate includes variances equal to zero. In this
case, the Gaussians degenerate into Dirac delta functions, and the model is
exactly equivalent to the dataset X itself: this is a useless solution. A solution
to avoid this problem consists of maximizing a penalized likelihood instead of
the likelihood itself; see [91].
As mentioned earlier, the EM algorithm is gradient based. In other words,
its convergence to the global maximum is not ensured in the general case. In
particular, assume that the likelihood admits several local maxima (that is,
points where dp{X\6)/dO = 0); then the solution found by the EM algorithm
is one local maximum, determined by the initialization point. Solutions to
overcome this problem consist of implementing other optimization methods
such as simulated annealing or stochastic EM [544], but none of them is simple
and/or computationally cheap.

2.2.9 Penalized Likelihood Approaches

Penalized likelihood approaches are really useful when a parameter estima-

tion problem admits many possible solutions, including trivial ones (problems
which lead to such trivial or useless solutions are said to be ill posed). In
this case, the ML estimate found is generally not the one sought. This can
however be avoided by modifying the likelihood into a penalized likelihood^
defined as
L ^ W = Lx(0) + Ar?(0), (2.47)
 2 An Introduction to Signal Processing 39

where Lx(^) is defined in (2.34) and f2[6) is aimed at lowering the likelihood
in parts of the space which are to be avoided. ^^ In the Gaussian mixture
example where J > m^ the penalty term can be set so as to disable variance
parameters that are too small; see [91]. The EM algorithm can be adapted in
order to address penalized likelihood problems [251].
An important application of penalized likelihood concerns the model se-
lection problem^ where one tries to estimate the probabilistic model that best
fits some data. A typical example is that of autoregressive models.

Autoregressive Models

Autoregressive models form a major signal processing tool. They are used
for spectrum estimation, coding, or noise reduction. An autoregressive (AR)
model (also called a linear prediction model) expresses a signal x{n) at time
n as a linear combination of its previous values:

x{n) aix{n 1) + a2x{n ~ 2) -\ + apx{n p) + e(n), (2.48)

where a = [ a i , . . . , ap]^ is the vector made of the AR coefficients. The order
of this model is the number p of coefficients. Assuming the noise e(n) is zero-
mean, white and Gaussian with known variance cr^, the log-likelihood of the
AR coefficients and the model order is
T
U{a,p) = - - log {27ra^) - y^Yl H^^^) " ^^'^^n-i-.n-pf. (2.49)
n=l

where -Kn-i.n-p = [x{n - 1),.. .x{n - p)]^ and assuming, e.g., Xn-i-.n-p =
[0,0,..., 0]^ for n < ]9. Here, we denote by x the vector made of signal sam-
ples x(n), n = 1,...,T. Maximizing Lx(a,p) yields uninteresting solutions,
typically p becomes very large. This problem can be overcome by maximizing
instead the penalized log-likelihood (2.47). Choosing Q{a,p) = p and A = 1
leads to Akaike's penalized log-likelihood, which leads to reliable model order
estimation [252]:
T
ir{a,p) = -|log(27ra)-^^||x(n)-aTxj;f-p. (2.50)
n=l

The next section presents an estimation framework different from likeli-

hood approaches: Bayesian statistical methods.

2.3 Bayesian Statistical Methods

In maximum likelihood parameter estimation, it is implicitly assumed that
there exists a fixed, ground truth parameter value. It can be argued, however,
^^See [252] for a review of penalized likelihood approaches.
40 Manuel Davy

that the parameter is well defined insofar as the model is correct, but in real
cases, the model is always an approximation of the real world. In Bayesian
approaches, it is assumed instead that the unknown parameter is a random
variable, characterized by a pdf. This pdf is stated before any data are col-
lected and is called the parameter prior distribution,

2.3.1 Bayes's Rule and Posterior Distributions

The most important concept of Bayesian statistical theory is the so-called

Bayes's rule. Consider two random variables ai G Ai and a2 G ^ 2 , with joint
pdf p(ai,a2). Recall (2.29): p(ai,a2) = p(ai|a2)p(a2) = p(a2|ai)p(ai). We
can deduce Bayes's rule

^o 1^ ^ p(ai|a2)p(a2) , .
p(a2 ai) = J -, (2.51)
J^^p(ai|a2)p(a2)(ia2
which enables us to 'reverse the conditioning'. Note that the denominator in
(2.51) equals / ^ p(ai, a2), daL2 = p(ai), but we keep the integral form in order
not to be confused with a possible prior over ai which would also be denoted
p(ai). Bayes's rule can be applied straightforwardly to Bayesian parameter
estimation. Assume we want to learn the value of some parameter 0 E 0 from
a set of data X. We already have the likeUhood p(X|0), and, being Bayesian,
a parameter prior p(0) is selected. Using Bayes's rule,

where p{0\X) is called the parameter posterior distribution^ which incorporates

information from both the data and the parameter prior. Roughly speaking,
the posterior provides information about the probability of each possible value
of 0 given the data and some prior knowledge. In practice, it suffices to write

p{e\X) a p(X|0)p(), (2.53)

where (x stands for 'proportional to'. The proportionality constant only de-
pends on the data X and it does not give additional information about 6.
Bayesian statistical methods are concerned with posterior distributions.

2.3.2 Bayesian Estimation

Similar to ML estimation, where estimation is performed from the likelihood,

it is possible to perform Bayesian estimation from the posterior. In particular,
two estimation methods are generally considered:
Maximum a posteriori (MAP). The estimated value maximizes the poste-
rior pdf
0MAP = argmax p(0|X).' (2.54)
eee
 2 An Introduction to Signal Processing 41

Minimum mean square error (MMSE). The estimate is the expectation of

the posterior pdf
0MMSE = ^p(0|X)[^]- (2.55)
An important remark is that Bayesian estimators can be 'biased': for ex-
ample, if the likelihood p(X|0) and the prior p(0) are Gaussian, with the prior
mean /x being different from the 'ground truth' parameter value 0gt, then
the above Bayes estimates are equal^^ (^MAP = ^MMSE) and their expectation
differs from Ogt. However, in the Bayesian philosophy, Ogt is meaningless as
there is no true value of 0 explicitly or implicitly assumed: all values 0 are
possible true values.

2.3.3 Bayesian Interpretation of Penalized Likelihood

In (2.47), the penalty term is introduced in the log likelihood. Going back to
the likelihood via the exponential function, we can rewrite it as

exp(LP""(0)) = p(X|^) xexp(Ar2((9)) (2.56)

and it appears that the likelihood has been multiplied by a term that only
depends on 6. Assuming that the following integral is finite,

0Lexp(AJ?(0)) dO = C, (2.57)

then we can define the pdf of 6:

P W = ^exp(Ar2(0)). (2.58)

We recognize in (2.58) the prior pdf of 9, and (2.56) can be interpreted

as the product of the likelihood with the prior, that is, the posterior p(^|X)
(up to a normalizing constant). Maximizing the penalized likelihood becomes
equivalent to performing MAP estimation.

2.3.4 Monte Carlo Methods

Bayes's theory provides a general framework for statistical inference. The main
concept is that of posterior distributions, which result from both the likelihood
and the prior. However, computing estimates such as (2.54) and (2.55) is
a difficult problem in the general case. For example, MMSE estimates are
obtained by the following integral:

0^ / ep{e\x)de. (2.59)
j0
^^It is worth mentioning that the product of two Gaussians is Gaussian; thus the
posterior pdf is also Gaussian in this example, which explains why its mean value
coincides with its maximum.
42 Manuel Davy

In statistical inference problems, the dimension of 0 may be very large (some-

times, more that 1000). Whenever the closed-form expression of (2.59) cannot
be found, a numerical integration technique should be implemented. Consider
the slightly more general integral computation case

7[h] = / h{e)7r{e)de, (2.6O)

Je
where 7T{6) denotes any pdf of 0, such as the posterior p(^|X). When the
dimension of is small (typically, smaller than 3), / [h] can be computed
numerically a la Riemann using a regular grid (provided 0 is a compact set).
For example, if 6 is one dimensional, and O [0,1], Riemann integration
consists of computing

1 ^
^[h] ^ -J2^{i/N)n{i/N). (2.61)

The limit of this approach is that the grid size increases exponentially with
the dimension de of 0 : assuming 100 grid points are used in each dimension
of of dimension do = 50, the grid size is 100^^ == 10^^^; this is out of reach
of today's computers.
However, another numerical computation technique can be implemented.
Assume random samples 6^^\ i = 1,...,A^ are available, where each sam-
ple 0^*^ is distributed according to 7T{0) (this is denoted ^^^^ ~ TT{6) in the
following). Then, from the law of large numbers,

^^t*^] = ^ E h ( 0 ( ' ) ) J[h]. (2.62)

The random samples 6^^\ i = 1,...,A^ are generally referred to as Monte

Carlo samples and IN [h] is the Monte Carlo estimate of / [h].
The estimate IN [h] is unbiased for any N, i.e., E^(^(i) 0(N)\(JN [h]) =
/ [h] and consistent. Finally, the empirical mean square error aj^ [h] provides a
reliable indication of the variance of IN [h] (in the sense that, asymptotically
as A/" ^ oo, a'fj [h] tends to the true estimation variance):

W - ^y VE [ h ( ^ ^ ^ V ^ i v [ h ] ] ' . (2.63)
2=1

Monte Carlo methods may also be used to compute other kinds of esti-
mates, inside or outside of the Bayesian framework. For example, Monte Carlo
optimization methods may be used to compute maximum likelihood of MAP
estimates; see [544].
We have assumed so far that Monte Carlo samples are available. The real
difficulty is actually in generating these samples. When 7r(0) is a standard pdf
 2 An Introduction to Signal Processing 43

(e.g., uniform or Gaussian), it is possible to implement direct sampling using

a random variable generator. In other cases, typically when n{9) is a posterior
of large dimension with an unknown normalizing constant, the problem can
be much harder. Several techniques have been developed for random variable
generation from any pdf. We present here Monte Carlo Markov chain (MCMC)
methods and importance sampling. ^"^

2.3.5 Monte Carlo Markov Chain (MCMC) Methods

The principle of MCMC algorithms is as follows: Given some pdf 7r{0) we want
to sample from, a chain of samples is generated iteratively; see Algorithm 2.2.
The chain is statistically fully determined by the pdf of the initial sample 0^^^
denoted 7ro{0) and a so-called Markov kernel IC{6\0'), which is a pdf w.r.t. 6
(for fixed 6'). Provided the kernel IC{6\0') satisfies some properties [544], the
pdf of each sample 6^'^^ slowly converges to the target pdf 7T{0) as i increases. In
particular, the kernel must be built so that 7r{0) is the invariance distribution
of/C(0|^'), namely.

L 0
lC{e\e')'K{e') dO' = niO) for all ee0. (2.64)

Algorithm 2.2: Generic MCMC Algorithm

1. Initialization: Sample 6^^^ -^ 7ro(^).
2. Iterations: For i = 1,2, .^., AT, do
Sample 5^^) - / C ( ( 9 | ^ ( ' - ' ^ ) .

Markov chain algorithms provide a series of samples which are asymptoti-

cally distributed according to 7T{0). In practice, it is necessary to run 'burn-in'
iterations before reaching convergence. Burn-in samples are not kept for the
computation of Monte Carlo estimates (2.62). Figure 2.9, p. 46, shows typical
random chains, including the burn-in iterations: the curves show more fiuctu-
ation of the current value in the beginning (during the burn-in) and then the
value converges.
In can be hard to design such a kernel from scratch. Fortunately, two sim-
ple algorithms enable easy building of such kernels: the Metropolis-Hastings
algorithm^ based on accept-reject moves, and the Gibbs sampler, based on
conditional sampling.

The Gibbs Sampler

Assume we want to generate Monte Carlo samples from p(^), with 0 =

[^1,..., 6do]^, where do is the dimension of the space 0. Assume moreover
13
See for example [544] for a complete overview of these techniques.
44 Manuel Davy

that we can sample easily from each of the conditional pdfs p(^i |^2, > ^de)^
p(6>2|<9i, (93,..., ^de). ^ Pi^do|<9i,..., Ode-i)' Typically, this situation hap-
pens when some of the conditionals are Gaussian and the others are, for
example, gamma distributions. The Gibbs sampler consists of samphng one
component of 0 at a time from the conditional posteriors, as presented in
Algorithm 2.3.

Algorithm 2.3: The Gibbs Sampler

1. Initialization: Sample 6^^^ ~ 7ro(^).
2. Iterations: For i = 1 , 2 , . . . , AT, and for j = 1,... ,de, do

. Samples^*'' ^ piOilOfK... ,J-'^J.

. Sample e^') ~ p{02\e^\ot'\-,0^J-'^).

. Sample?*^ ~ p(^del^i''. >^!i^-i)-

The Gibbs sampler is quite simple; however, it requires the ability to sam-
ple from the conditional pdfs. This is sometimes not possible to implement,
and we can use instead the Metropolis-Hastings (MH) algorithm.

The Metropolis-Hastings Algorithm

In addition to the target pdf 7r(0), the MH algorithm requires a proposal pdf
q(0|') that we can directly sample from, such that q{0\0') ^ 0 whenever
7r(0) ^ 0. Algorithm 2.4 presents its structure.

Algorithm 2.4: The Metropolis-Hastings Algorithm

1. Initialization: Sample 0^^^ ~ 7ro(^).
2. Iterations: For i = 1 , 2 , . . . , AT do

Sample a candidate parameter value 0* ~ q{6\6^^~^^).

Compute

a^H{e\0'<'-'^) = min (2.65)

7r((9(^-i)) g(r|(9(^-i))

With probability QMHC^*,^^*"^^), accept the candidate, i.e., set 6^'^ ^ ^*.
Otherwise (that is, with probability 1 - aMH(^*,^^*~^^), reject the candidate,
i.e., set(*) ^0^'-^l

From Algorithm 2.4, it is clear than the MH kernel /C(0|0') is determined

by the proposal pdf q{O\0^). Note that the pdf 7r{6) appears in a ratio: it can
be known only up to a normalizing constant (which is often the case for a
posterior pdf due to Bayes's theorem) because it cancels out. Various choices
of proposals lead to various MH algorithms. Three important subcases are
presented below.
 2 An Introduction to Signal Processing 45

Set q{O\0^) = Af{0; 0\ U). This corresponds to a random walk proposal

It also is a symmetric proposal in the sense that q{0\6') = q{0'\6), and
the ratio in (2.65) does not depend on q. This kind of proposal is often
called a local proposal because the chain evolves locally around the last
accepted point. Such chains have the advantage of not remaining at the
same point for many iterations: the fraction of accepted candidates can
be set to about 50% by tuning the covariance matrix U. However, local
proposals cannot explore the whole space 0 very quickly.
Set q{0\O') = q(0), independent of 6'. This proposal is called independent
or global As opposed to local proposals, it does not consider the last ac-
cepted value when building the candidate. This permits large jumps over
the space 0 , and quick convergence. However, it also causes many steady
points: the chain may keep at the same value for many iterations (i.e.,
g(i) ^ g(i+i) :=,.., = Qii^i) ^here / typically equals 100 or 1000) because
all the proposed candidates are rejected. It is thus quite important to build
global proposals upon heuristics about the regions of 0 where 7r(0) is likely
to be large.
For each component Oj {j = l , . . . , d e ) , use a proposal q{0j\6j) while
not touching the other components. Overall, this proposal can be written
qj{e\e') = qiOjie^^) nt=i,i^j ^^;(^^) for any j = 1 , . . . , de^ where Su{v)
is the Dirac delta function.^^ This is a one-at-a-time proposal pc?/because
it only updates one component of 0 at each iteration, using a local or a
global proposal.
Figure 2.9 displays example outcomes for the proposal distributions listed
above. These MH kernels can be mixed, yielding another admissible ker-
nel: Given a family of MH kernels {ICj{0\0'),j = 1 , . . . , J } that have the
same invariant pdf 7r(0), and positive coefficients {/?j, j = 1 , . . . , J } such that
J2j=i 03 ^^ ih.en the mixture kernel

K{e\e') = Y^fij^Mo') (2.66)

also has TT{0) as invariant pdf. In practice, this kernel is implemented as

follows: At each iteration, sample a discrete variable u in { 1 , 2 , . . . , J } with
discrete distribution P[u = j) = /3j^ and use the kernel with the index j = u
drawn. This leads to, e.g., kernels that update all components one at a time
(using a mixture of kernels with one-at-a-time proposals for each component),
and mix local and global moves for faster convergence; see Fig. 2.9.

^^The Dirac delta function can be viewed as the derivative of the step function
at point 0, where the step function equals zero over ] oo,0[ and one over ]0, oo[.
An important property is that Su{v) = 0 whenever u ^ v and for a function h(i;),
we have J \\(v)6u(y)dv = \\{u). When used in a probabilistic context, writing a has
distribution (5n(a) means that a. = u deterministically.
46 Manuel Davy

o.sf 0.5L 0.5

0.25 0.25| 0.25

1000 0 1000 1000

a) Local proposal b) Global proposal c) Mixture proposal

F i g . 2.9. Typical Markov chains produced by different Markov kernels for the
Metropolis-Hastings algorithm. The parameter sampled is a frequency having the
true value 0.25. a) Local, Gaussian random walk proposal pdf. b) Global, indepen-
dent proposal pdf. c) Mixture of a local and a global proposal. The local proposal
fails to explore the frequencies in a short time, whereas the global proposal keeps at
the same frequency for many iterations. The mixed proposal performs well.

A n E x a m p l e : B a y e s i a n E s t i m a t i o n of S i n u s o i d s i n N o i s e

In order to examplify the use of M C M C methods, we present here a simplified

version of the M C M C algorithm of Andrieu and Doucet [21] which is aimed
at estimating the parameters of an unknown number of sinusoids in noise,
as well as the number of sinusoids itself. Here, we consider an extension of
the model (2.31) with M sinusoids in Gaussian noise, with M known. The
unknown parameters are the frequencies (denoted k = [/ci,..., kM])^ the am-
plitudes, the initial phases, and the Gaussian noise variance cr^. This model
can equivalently be written as
M
x{n) = ^ a^ cos{27rkmn) + o:^ sin(27rA:mn) + e(n) for n = 1 , . . . , T,
m=l
(2.67)
where the (non-linear) phase parameters have been replaced by an additional
set of (linear) amplitude parameters. Let us define a = [af, a ^ , . . . , a%j, Q^MI^
and
cos(27rA:i 1), cos(27rA:i 2), cos(27rA:i T )
sin(27r/ci 1), sin(27r/ci2). sin(27rA:i T )
D(k) = (2.68)
cos(27r/cM 1), cos(27rfcM 2), . . . , cos(27r/cM T)
sm{27TkM 1), sin(27r/cM 2), . . . , sin(27rA:jvf T)
Then, using the notations introduced in Section 2.2.5, p. 32, (2.67) becomes

X = D ( k ) a + 6, (2.69)

corresponding to t h e Gaussian likelihood p(x|k, a,cr^) = A r ( x ; D ( k ) a , c r ^ l ) .

In order to estimate the unknown parameters, we can embed this model into
t h e Bayesian framework by defining parameter prior pdfs, as follows:
 2 An Introduction to Signal Processing 47

The amplitudes follow a Gaussian prior pdf with mean 0 and covariance
a'^^cti which is proportional to the additive noise variance (this is to adjust
the noise 'amplitude' to the sinusoids' average amplitude).
p(a) - A/'(a;[0,0]T,a2i;), (2.70)

where U^ is a covariance matrix chosen as Ea = 7^ [D^(k)D(k)] (this

is the g-prior, see [21]). The parameter 7^ can be interpreted as an expected
signal-to-noise ratio, and it is thus quite difficult to set its value a priori.
Consider now 7^ as a unknown random hyperparameter to be estimated,
and set its prior p(7^) = ^ ^ ( 7 ^ ; 1^2, ^^3); see (2.72). Setting 1^2 = 2 ensures
that p(7^) has an infinite variance (it is non-informative), and the value
of u^ has very limited influence (1/3 = 20 is the standard choice).
The frequencies may be anywhere in the interval [0; 0.5], and thus we select
a uniform prior pdf
M
p(k) = n ^0,0.5l(fcm)- (2.71)
The additive noise variance should be as small as possible, and we select
its prior as an inverse Gamma distribution

(2.72)
where r{-) is the Gamma function [8]. This choice has two main justifi-
cations: First, for small i/Q and i/i, this density favors small values of cr^.
Second, this pdf is called a conjugate prior because the posterior distribu-
tion can be calculated in closed form. In practice, we may choose i^o <^ 1
and 1/1 <C 1 (e.g., I/Q = ^1 =^ 10~^); the precise selected values have little
influence on the estimation results.
Using these priors and the likelihood, it is possible to write the para-
meters posterior. Moreover, we can express the two conditional posteriors
p(k, a , cr^|7^,x) and p(7^|k, a , cr^, x). The former can be further decomposed
into the product p(k|7^,x)p(a^|k, 7^,x)p(a|k, cr^,7^,x), with

p(k|7^x) = - ^ [^o + x T p ( k , 7 > ] ' ^ ^ , (2.73)

p(<T2|k,7^x) = Xe ( ^ ^ 2 . I + i l ^ ^ ^ + ^ T p ( k , ^ 2 ) ^ ^ ^ (2.74)

p(a|k,(T2,7',x) = A/-(a;S(k,7')D(k)'rx,a2S(k,7')), (2.75)

whereP(k,7') = lT-D(k)S(k,72)D(k)TandS(k,72)=j^[D(k)"rD(k)]-\
The hyperparameter conditional posterior is the inverted gamma pdf

p(7^|k,a,cT^x) = j g ( 7 2 ; ^ 3 + i , ^ ! H ( ^ M ^ + ^ , ^ , (2.76)
48 Manuel Davy

where the parameters are chosen as z/3 = 2 and z/4 = 20 to ensure t h a t the
prior over 7 has an infinite variance and a vague shape. Following [21], we can
now write the M C M C algorithm t h a t mixes a Gibbs sampler with one local
and one global MH kernel.

Algorithm 2.5: A n M C M C Algorithm for Sinusoids in Noise Estimation

1. Initialization:
Sample 7^^^ ~ Z6?(72; 1/3,^^4).
For m = 1 , . . . , M, sample/cm ~ q^(A:m|x).
2. Iterations: For z = 1, 2 , . . . , AT do

For m = 1 , . . . , M do
- Frequency MH step
With probability /3, sample a candidate using a local proposal k^ ~
^l\'^m\^Cm )-
Otherwise, sample a candidate frequency using a global proposal k^ ~

Compute QMHC^*,^^'"^^) using (2.65) with target pdf p(fcm|7^^*~^\x)

and with the proposal selected randomly above, ^hen perform the ac-
cept/reject step (see Algorithm 2.4). This yields /cm
- Noise variance and amplitude direct sampling
Sample the noise variance a^^*^ using p(cr^|k^'\7^^'~^\x) given in (2.74).
Sample the amplitudes a^*^ using p ( a | k ^ ' \ a ^ ^ ' \ 7 ^ ^ ' ~ ^ \ x ) given in (2.75).
- Hyperparameter sampling
Sample 7^^^^ using p{'y'^\k^'\a'^'\a'^^'\x) given in (2.76).

In Algorithm 2.5, the frequency parameter is sampled using a mixture of

two M H kernels t h a t rely on either a local proposal, chosen as q/(A:m|^m ) =
N{krn'-,i^m 5^RW)' whcrc (j\^ is a user-defined r a n d o m walk variance, or
a global pdf chosen proportional to the frequency spectrum^^ of x, i.e.,
q^(fc^|x) oc \X{km)\^- T h e mixture coefficient is selected as /? = 0.75 (stan-
dard choice). After the frequency is sampled, the noise variance parameter and
amplitudes are sampled directly from t h e conditional posterior pdfs. These
three sampling operations produce a sample (k^*\ a^*\ 5^^^^) at each iteration
i. This is the first step of the overall Gibbs sampler; the second step consists
of sampling the hyperparameter 7^^*\ Figure 2.10 displays an example chain
built using this algorithm. This algorithm may be complex for estimating one
or two sinusoids; however, its settings remain the same for large number of
sinusoids, which makes it an attractive solution in such cases; see Chapter 7.

^^It is easy to build a random variable generator from a stepwise approximation

of this pdf (onto k{j), j = 1 , . . . , J ) by first sampling a continuous uniform ran-
dom variable u on [0,1], then choosing jo such that (1/C) Y.%~i^ l ^ ( ^ ( i ) ) P < '^ <
( l / C ) E ^ i i l^iHM'^. where C = S / = i l ^ ( ^ ) l ^ and finally sampling k uniformly
on [k{jo-l),k{jo)].
 2 An Introduction to Signal Processing 49

0.25

"''rTiTf 1 'I yi 1 1 n

1000 0 1000 1000

Frequency ki Amplitudes af and al Noise variance cr^

Fig. 2.10. The frequency k and amplitudes a vectors and the noise variance a^
sampled by Algorithm 2.5 for m = 1 and M = 12. As can be seen, convergence
is reached after 200 iterations. The true parameter values are mi = 0.12, a\ = 1,
al = 0, and a^ = 2.

Algorithm 2.5 can be extended to an unknown number of sinusoids in

noise; see [21]. However, considering the number of sinusoids as a parameter
to be estimated makes the implementation a bit more complicated, though
still feasible. In particular, the parameter space dimension changes whenever
the number of sinusoids changes, and a special algorithmic structure must be
used, such as the reversible jump MCMC algorithm, which includes sinusoid
birth and death moves [20].

2.3.6 I m p o r t a n c e S a m p l i n g a n d S e q u e n t i a l I m p o r t a n c e S a m p l i n g

Another interesting technique for the computation of Monte Carlo integrals

is importance sampling. Let 7T{6) be some pdf, for example a posterior pdf in
the Bayesian setting. Let q{0) be another pdf (this is a proposal pdf, called
here an importance pdf) such t h a t q(0) 7^ 0 whenever 7r{G) ^ 0. Moreover,
q(^) is selected so t h a t it is possible to sample directly from it. T h e n (2.60)
can be written

/[h] = l^h{e)7,{e)de = ^ h w ^ q ^ d . (2.77)

Now, assume a set of Monte Carlo samples ^^^\ i = 1 , . . . , TV are generated

using q{6). Then the following Monte Carlo estimate holds:

N
/^[h] = ^5(^)h(6l^^)) ^ /[h], with 5 (i) (2.78)
i=l
q(0(^))'

where i}(^) is the importance weight of sample ^^^^, and is aimed at correcting
the discrepancy between q{0) and 7r(0). A key remark is t h a t the variance of
the estimator (2.78) strongly depends on the importance pdf q(^) selected. It
can be demonstrated t h a t , for a given number of Monte Carlo samples, the
50 Manuel Davy

optimal importance pdf t h a t minimizes the variance of the estimate (2.78)

is q(0) oc \h{6)\7T{6). However, it cannot be applied because it is usually
impossible to sample directly from it, and the normalizing constant is hard
to compute. ^^ Nevertheless it provides some hints for the selection of q(0): it
should be as close as possible to the optimal importance pdf.

Particle Filtering

T h e main application of importance sampling is particle filtering. Assume

a dynamical system with hidden state parameter On and observation x ^ for
n = 1 , 2 , . . . is governed by the following equations^ "^ (see also the graphical
representation in Fig. 2.11):

On = fn[On-i] + v^ (transition equation), (2.79)

x ^ = 5'n[^n] + ^n (obscrvatiou equation). (2.80)

/ n + l [ J + Vn-f 1
^^

gn-l['] +^n-l \gn[']+^n

fXn-1j

Fig. 2.11. Graphical illustration of the dynamic model in (2.79) and (2.80). In this
model, only Xn is observed. The state parameter vector On is hidden for n = 1,2,
On follows a Markov evolution, hence the name hidden Markov model.

Assuming the pdfs of v^ and n are known, we can express (2.79)-(2.80) as

the state transition prior p ( ^ n | ^ n - i ) and t h e observation likelihood p(xn|^n)-
Together with the initial state distribution po{Oo)^ this defines a sequential
Bayesian model. Such dynamic systems are often met in practice: for example,
On is a vector composed of the position, speed, and acceleration of a plane at
time n, and Xn is the observation received by radar. T h e problem is t h a t On
cannot be observed directly; it has to be estimated.

^^Whenever h is always positive, the normalizing constant is / [h] itself!

^^For consistency with most works in the field, and for notational simplicity, we
use subscript time indices in this subsection, e.g., On is the state at time n, instead
of the usual notation 0{n).
 2 An Introduction to Signal Processing 51

The overall objective of filtering is the sequential estimation of the state

trajectory for times between 0 and n, denoted 9o:n {^o, ^ i , ? ^n} via the
posterior p(0o:n|xi:n) at each time n. In the special case where the functions
fn[-] and gn['] are linear, and the noises v^ and Sn are Gaussian, this can
be solved analytically: the posterior is Gaussian, and its mean and covariance
matrix are provided by the Kalman filter [18]. In the general case, the esti-
mation of Oo:n relies on intractable integrals computation (such as (2.55) for
MMSE estimates) at each time n. Particle filtering provides a Monte Carlo
approximation via sequential importance sampling, where p(0o:n-i|xi:n-i)
plays the role of 7r(0) in (2.77). In order to make the problem sequential, we
first write the posterior p(0o:n|xi:n) at time n as a function of p(0o:n-i|xi:n-i)
at time n 1:

rn \ \ (Ck I X P(^n|^n-l)p(Xn|^n) /ooi\

P(^0:n|xi:n) = p(^0:n-l |xi:n-l) ^ 1 \ T , (2.81)
p(Xn|Xi:n-lj
where p(x^|xi:n-i) is a normalization term which need not be computed.
Second, we select an importance pdf that can be written sequentially, namely
n
(^0:n-l)qn(6>n|^n-l) = ^^(0^)^^l{H^l-l)^ (2.82)
Z=l

It is now possible to sample the state at time n using qn(^n|^n 1) and com-
pute sequentially the weight as the ratio p(0o:n|xi:n)/qn(^O:n)- These elements
lead to the particle filter presented in Algorithm 2.6 below.

Algorithm 2.6: Particle Filtering Algorithm

1. Initialization: For each particle z = 1,... ,N, sample independently OQ ~ qo(^o)

and compute the initial weight UJQ po(^o )/qo(^o )
2. Iterations: F o r n == 1, 2 , . . . , AT do

The particle trajectories are updated using the sequential importance pdf
- For i = 1,... ,N, sample the new state at time n for particle i,

0W ~ q n ( 0 n | 0 i ^ l j . (2.83)

The weights are computed and normalized

- For i = 1 , . . . , A^, compute the sequential importance weight ujn as follows:

qn[tfn \t^n-l)

- Compute the weight normalization constant Wn = ^i=i <^n (this normal-

izes the weights, and avoids computing the term p ( x n | x i : n - i ) in (2.77)).
- For particles i = 1,... ,N, set UJH^ ^ uJ^n^/Wn-
The current state in estimated
- Estimate the state from the weighted particles, e.g., ^nMMSE = = S ^ i ^ n ^ n
52 Manuel Davy

The particles are resampled

- Compute the efficiency number iVeff = [Y^iLii^^^)^]~
- If A^efF < A^hreshoid, then resample the particles: duplicate particles with large
weights, suppress particles with small weights, and set Un < l/N for
i = l,...,iV.

At each time n, Algorithm 2.6 produces Monte Carlo samples oVn' ^

1 , . . . , A/' (called particles in this context) with weights uJn which approximate
the posterior p(0o:n|xi:n). However, this simple strategy produces samples
which degenerate: after a few iterations, the weights of most particles are
close to zero, whereas the weights of a few particles are significantly non-zero.
This can be monitored by A^eff' if it becomes too small, then degeneracy has
occurred and particles are resampled; particles with low weights are moved to
more accurate points (i.e., to the location of particles with large weights). This
can be implemented by first setting randomly the number of copies of each
particles using, e.g., stratified samphng [344], where the expected number of
copies of each particles equals its weight times A^. Particles with 0 copies are
suppressed.
Different importance pdfs lead to different algorithms. A simple choice
consists of using the transition pdf qn(^n|^n 1) = p(^n|^n !) The
corresponding algorithm is simple but the variance of estimators based on the
particles generated is large in general. The importance pdf that minimizes
the variance of the weights (and thus, the variance of estimators) uses the
new observation. It is written qn{0n\0n 1) = p(^n|xn,^n-i), but it is
often impossible to sample from it. However, it can be approximated locally
by Gaussians: qn{0n\0n - 1) = A/'(0n;^n|n, ^ n | n ) where Onin and Un\n are
the state and state covariance estimates given by the extended Kalman filter
or by the unscented Kalman filter. ^^

2.4 Pattern Recognition Methods

Pattern recognition algorithms are mainly concerned with data classifica-
tion.^^ There are generally two frameworks considered. In the supervised clas-
sification framework, a set of labelled training data is provided. A set of pairs
^^The Kalman filter is devoted to linear and Gaussian dynamic models. When
the model is non-linear and/or non-Gaussian, the posterior is non-Gaussian but it
is still possible to approximate it by a Gaussian. This can be implemented by the
extended Kalman filter, which operates on linearized state equations (2.79)-(2.80)
with Gaussian noises, or by the unscented Kalman filter, which approximates the
posterior as a Gaussian, without linearizing (2.79)-(2.80).
^^Detection is also a pattern recognition problem. However, it can be consid-
ered, to some extent, as a subcase of classification. For the sake of simplicity, here
we mainly cover classification. Some elements about detection theory will also be
provided when appropriate.
 2 An Introduction to Signal Processing 53

(X,Y) = {(xi,yi),...,(x,n,ym)} is given, where x^ G A* and y^ G ^ for

i = 1 , . . . , m. Example data and label spaces are X = R^^ and 3^ = { 1 , 2 , . . . , }.
Labels indicate to which class each data belongs. The aim of supervised clas-
sification is to predict the class label y G 3^ of a new datum x G A' which is
not in (X, Y). This prediction must be elaborated from the sole knowledge of
X and of the training set (X, Y). In the unsupervised classification framework
(which is often called clustering), a set of unlabelled data X = { x i , . . . , x ^ }
is provided, and each datum in X must be assigned a label from 3^ (i.e, it must
be assigned to a class).
In this section, we review classification techniques. One unsupervised clas-
sification method has already been presented in Section 2.2.8, p. 35. Indeed,
Gaussian mixtures models assign a latent variable Zi to each datum x^, which
indicates to which Gaussian x^ is related, that is, to which Gaussian class it
is related. The number of classes J need not be known, because a penalized
likelihood approach [91] or a Bayesian approach [541] could be implemented
to estimate both the labels and the number J of mixture components (that is,
the number of classes) found. Another unsupervised classification technique is
presented in the next subsection, while the remaining subsections are devoted
to supervised classification.

2.4.1 K-Means
The i^T-means algorithm is an unsupervised classification method. The set
of data X = { x i , . . . ,Xm} in X is provided. Given the number of classes,
denoted K, the algorithm assigns a label in 3^ = { 1 , . . . , K} to each x^. The
algorithm is summarized below, where d(-, ) is a distance measure in X,

Algorithm 2.7: /C-Means Algorithm

1. Initiaiization: Set the K Initial means /x^ m X, j = 1,... ,K. For example, an easy
and robust choice consists of choosing any K (not equal) data in X. This ensures
that no cluster is empty in the beginning.
2. Iterations: While the means keep evolving, do
For each datum Xi, i = 1 , . . . , m,
- Compute its distance to each means d ( x i , /x^), for j = 1 , . . . , K.
- Assign to Xj (i = 1 , . . . , m ) the label y^ = argmin d(xi,/x ).
3= 1,...,K
Update the means by setting
/x^ ^ mean{xi|yi = j } , j = \,...,K. (2.85)

This algorithm converges to useful solutions insofar as the data in X form

separated classes. Of course, the classification results are closely related to
the distance measure used in Algorithm 2.7. Typical distance measures are
the Euclidean distance d(xi,X2) = (xi X2)^(xi X2) = ||xi X2||^, the
Li distance d(xi,X2) = ||xi X2||A' or the Mahalanobis distance d(xi,X2) =
(xi X2)^X'~^(xi X2), where E is the empirical covariance matrix of the
set X, computed as in (2.86) below.
54 Manuel Davy

2.4.2 Principal Component Analysis

In front of an unlabelled dataset X in the space X^ one may want to reduce

the data dimension. This may be because this dimension is too large to en-
able the use of a classification technique (e.g., Gaussian mixture model). It
may also be to enable 2D data visualization of larger dimensional data. Prin-
cipal component analysis (PCA) is aimed at reducing the dimension of data
without losing too much information. The amount of information carried by
a dimension is summarized in the empirical covariance matrix

^ r = ^Yl^^i-t^Di-^i-i^rv wither = -E^^' (^.se)

where /x^"^^ is the empirical mean of X.^^ In addition, S^"^ also informs us
about the redundancy between the dimensions. In order to reduce the number
of dimensions of the data, PCA provides us with the most informative linear
combinations of the dimensions of the data [161]. The steps of PCA are
1. Compute the empirical covariance matrix S^"^ as in (2.86);
2. Apply matrix diagonihzation to X'x"^^, i.e., compute
i;^^ = SAS-\ (2.87)
where A is a diagonal matrix whose diagonal terms are called the eigen-
values of E^"^. The eigenvalues are positive, and they are assumed sorted
in decreasing order along the diagonal. The columns of S are called the
eigenvectors^^ of X'x"^^, and the matrix S is orthogonal, meaning that

3. Keep the dpca columns of S that correspond to the dpca largest eigenvalues
in A. They are stored in a matrix Spca
with size dx x dpca and the
corresponding eigenvalues are in the squared matrix Apca with size dpca-
4. Compute the lower dimensional data for i = 1 , . . . , m:
C, = Spca(xi - Mx"")- (2-88)
The matrix diagonalization in (2.87) is almost always possible^^ whenever
m is larger than dx- The transform in (2.88) reduces the dimension of x^
^^The empirical covariance matrix in (2.86) is computed with normalization term
1/m. The corresponding empirical covariance matrix is known to be a biased esti-
mate of the true covariance matrix. The unbiased empirical covariance matrix should
be computed with normalization term l/(m 1).
^^ Details about eigenvalues, eigenvectors, and matrix diagonalization may be
found in any textbook about matrices; see, e.g., [295].
^^A property is said to be almost true (here, the property is 'diagonalizing the
covariance matrix is possible') if it is true with probability one. Surprisingly, even
though it is true with probability one, it is still possible that the property is false,
but this is quite unlikely; see [548] for further details. Here, the property would be
false if at least one datum was a linear combination of others.
 2 An Introduction to Signal Processing 55

{i = l , . . . , m ) , i.e., the dimension of (^^ equals dpca- The variables Ci have

covariance matrix Apca-
The PCA approach has two major advantages. First, it reduces the dimen-
sion of the data to the most informative dimensions. Second, it removes the
correlations between the initial data dimensions, which removes redundancy.

2.4.3 Supervised Classification: Cost Function and Risk

Supervised classification is concerned with the design of a classification func-

tion F(x,Y) ' ^ ^^ y which associates a class label y G 3^ to each datum x G A*,
and which is designed from the training set (X, Y). In this section, we restrict
to two-class classification and set 3^ = {1; 1}. Multi-class classification algo-
rithms can be derived from two-class algorithms; see [568].
Loosely speaking, the function F(X,Y) is relevant only if it has the smallest
possible risk of performing classification errors. Let F be any classification
function. Errors occur whenever F(x) = y', whereas the true class label of x
is y, with y' 7^ y. In practice, the seriousness of errors needs be measured in
order to design F(X,Y); this is the role of loss functions. Let c(x,y; F(x)) be
such a loss function for a classification function F. A standard example is the
0-1 loss c(x, y;F(x)) = 1 5y(F(x)). In the two-class classification case, it
is helpful to assume that F delivers values over R (and not just y ) , and set
y = sign (F(x)). This enables the use of, e.g., the quadratic loss, the hinge
loss, etc. (see Fig. 2.12).

c(x,y;F(x))
Quadratic

Hinge (y = + 1 ) ^ y Hinge (y = - 1 )
0-1 loss

F(x)-y
Fig. 2.12. Standard loss functions to be used in learning problems. The quadratic
loss c(x,y; F(x)) = (y F(x)) is used in so-called least square methods, whereas
the hinge losses are used in support vector machines. The 0-1 loss c(x,y; F(x)) =
1 (5y(F(x)) is often used in Bayesian classification problems.

From loss functions, we can define the risk R [F] of a classification function
F as the expected loss over all possible pairs (x, y), namely

R[F] = Ep(,,y)[c(x,y;F(x))], (2.89)

56 Manuel Davy

where p(x, y) is the 'pdf' of the data and labels. Of course, this 'pdf' is un-
known in practice, and the risk cannot be computed directly. Assume however
that R [F] of any classification function F could be computed: the optimal clas-
sification function could be found by minimizing the risk w.r.t. F over a set
of functions denoted T. This approach is not really possible, but the risk can
still be estimated from the training set. Define the empirical risk
^ m
Remp
(X,Y)
[F] = - V c ( x , , y , ; F ( x i ) ) (2.90)

as an estimate of the true risk R [F], based on the information provided by

the training set. A well-known problem is that R/xV) I'^l ^^ ^ P^^^ estimator
of R[F], and there usually exist many functions F such that R(xV) [^1 ~ ^'
including trivial ones. In Fig. 2.13, a classification function which achieves
a zero empirical risk is depicted. It fits the training data perfectly, but its
shape is so complicated that it is likely to make many errors on data which
are not in the training set. In other words, this function fails at generalizing
its correct classification ability to new data. Such a solution is said to overfit
the training set, and over-fitting is often related to generalization issues. As
explained in the next section, it is necessary to prefer solutions which are
simple in some sense, so as to limit over-fitting and favour good generaliza-
tion.

/^^i: 'ix^s^B-fu^.-^^,
D D

" 1 /
I'

Fig. 2.13. Supervised classification into two classes with 2-dimensional data. In the
training set (X, Y), data with label y = 1 are represented with dots, whereas data
with label y = 1 are represented with squares. The dotted line is a classification
function F such that R(x Y) ^\ ~ ^- Though it achieves zero empirical risk, F is not
a good classification function, as it makes an error for a new datum which is not in
the training set (circle at the bottom, with the true label y = 1).
 2 An Introduction to Signal Processing 57

2.4.4 Regularization

Minimizing the empirical risk leads to useless functions F. This problem can be
overcome by minimizing an objective function that includes both the empirical
risk R/xV) [^] ^^^ ^ term i? ( F ) that penaUzes unwanted solutions. Define the
regularized risk
R^^[F] - R^5^;V)[''] + ^ ^ ( F ) , (2.91)
where A is the regularization tuning parameter. The optimal classification
function F(X,Y) is found by solving

F(x,Y) = argmin R- [F]. (2.92)

FGJF

Given a training set, various supervised classification methods employ dif-

ferent loss functions, penalty terms and sets of functions ^ so as to design
F(x,Y)- Examples are penalized likelihood approaches (see Section 2.2.9, p. 38)
and Bayesian approaches (see Section 2.3, p. 39). The following subsection
presents another one: the support vector machine.

2.4.5 Support Vector Machines

Support vector machines (SVMs) are specific instances of the above regular-
ization scheme. In SVMs,

jr = {F(.) = f(.) -f b such that f G W and 6 R}, (2.93)

where W is a space of functions called a reproducing kernel Hilbert space with

kernel k(-,-).^^ In practice, the dot product defined over W, denoted (,-)'H?
can generally not be written in an explicit formall we need to know is that
it has the properties of a dot product, and that it enables the reproducing
kernel property for some kernel. In (2.93), the set T of possible classification
functions is specified by selecting the kernel k(-,-). Provided the kernel is
positive definite,^^ it gives rise to a RKHS. A common choice is the Gaussian
kernel (for any x and x' in A*)

^^A Hilbert Space 7^ is a nonempty space such that 1) 7i is a vector (linear)

space, that is, any linear combinations of elements of 7i are also elements of H; and
2) A dot product (, )n is defined over H. This dot product induces a norm for any
f e 7i such that ||f(-)ll?t = {H')^^{'))n 3) 7i is complete for this norm. Moreover,
assuming 7^ is a Hilbert space of functions f(-) : A' ^- M, it is a reproducing kernel
Hilbert space (RKHS) if and only if there exists a kernel function k(,): A* x A' -^ M
with the following reproducing property: For any x. e X, k(x, ) is in 7i and for any
function f G 7^, we have (k(x, ),f(-))'K = f(x).
^^A kernel k(-, ) is positive definite if and only if for any set {xi,... ,Xm} and
any m > 0, the matrix with entries k(xi,Xj), {i,j) = 1,... ,m is a positive definite
matrix.
58 Manuel Davy

k(x,x')=exp|-2i5l|x-x'f}. (2.94)

In SVMs, the hinge loss Chinge (x,y; F(x)) (see Fig. 2.12) is chosen and
the penalty term i^ (F) is the squared norm induced by the dot prod-
uct in W, i.e., i? (P) = ||f ||f^. For any x, from the reproducing property,
F(x) = (k(x, ),f(-))'^ -f 6, which means that classifying a datum x is an
affine operation (that is, a linear -h constant operation) in terms of elements
of 7i, because the dot product (f(*)?f'(*))w is linear w.r.t. both f(-) and f (),
and a non-hnear operation in terms of elements of A'. In SVMs, the regularized
risk (2.92) becomes
^ m
f^"[''] = - J ] c , , , . ( x y , ; ( k ( x , , - ) , f ( - ) ) w + b ) + A | | f | | 2 , . (2.95)
2=1

It is possible to modify this regularized risk slightly so as to replace the reg-

ularization tradeoff parameter A by another parameter v G [0,1] that can
be interpreted more easily. Minimizing this modified regularized risk can be
equivalently written as the soft margin 5VM optimization: Minimize

1 \ ^
-^\\\i-'^p-^ m 5. Z ^ i with respect to f,p,^i,b, (2.96a)
1=1

with

yi((k(xi,.),f(.))7i+&) >p-^i forall i = l , . . . , m , (2.96b)

and
Ci > 0, for alH = 1 , . . . , m , p > 0, (2.96c)
where the slack variables ^i {i = 1 , . . . , m) are used to implement the hinge
loss [568]: they are non-zero only if y^ ((k(x^, ')j{'))n + &) < P and, in this
case, they induce a linear cost in the objective function ^ ||f ||?{^P+^ S I ^ i ^i-
Introducing Lagrange multipliers a i , . . . ,am, the optimization in (2.96) can
be turned into the equivalent quadratic convex problem with linear constraints
(dual problem): Maximize
^ m m
- 2^ 2^ Oiiajyiyjk{xi,:x.j) with respect to a^ (z = 1 , . . . , m), (2.97a)
2=1 j = l

with

0 < Oil < 1 / ^ (i = 1 , . . . , m) , Y^ ^iVi 0 a^d 2_\ ^^ ^

2=1 2=1
(2.97b)
 2 An Introduction to Signal Processing 59

The set of Lagrange multipliers a^, 2 = 1 , . . . , m that solve (2.97) leads to the
optimal classification function F(X,Y)^ which is written for all x G A'
771

F(x,Y)(x) = 5 ] y , a i k ( x , , x ) + 6 . (2.98)

The classification of a new datum x into one of the two classes {1;+!}
is performed by assigning to x the class label y = sign (F(X,Y)(X)) where
F(X,Y) is given in (2.98). The support vector machine admits a gemoetrical
interpretation; see Fig. 2.14.

Fig. 2.14. Geometrical interpretation of the i^-soft margin support vector machine
in the reproducing kernel Hilbert space H. Each element of Ti is a function which
is plotted here as a point in an infinite-dimensional space (here, 7i is represented
as a two-dimensional space). The classification function f is the vector othogonal to
the hyperplane with equation (f, g)?t -\- b = 0. This hyperplane separates the data
from the two classes with maximum margin, where the dots represent the functions
k(xi,) with label y^ = 1, and the squares represent the functions k(xj, ) with
label yj = I. The margin width is p/||f||7^. As this is the soft margin SVM, some
training vectors are allowed to be located inside the margin. The vectors located on
the margin hyperplanes (dotted lines) and inside the margin are called the support
vectors.

An important property of this SVM is its sparseness: v is an upper bound

on the fraction of non-zero Lagrange multiphers. In other words, assume for
example v = 0.2; then at least 80% of the Lagrange multipliers {ai^i =
1 , . . . , m} equal zero, and computing F(X,Y)(X) requires us to compute k(xi, x)
for only a small fraction (at most 20%) of the training data x^, i = 1 , . . . , m.
Training data with non-zero Lagrange multipliers are called support vectors.
The support vectors such that 0 < a^ < 1/m are called the margin support
vectors, and we denote them by x^*. The set of margin support vectors is
60 Manuel Davy

denoted X* and it includes m* vectors. In (2.98), b is computed as the average

over all the support vectors:
-. m

^ = -;;^ E T.^'^yM^J'^i^)- (2-99)

2.4.6 Nearest Neighbours Classification

Nearest neighbours classification is a very simple supervised classification tech-

nique. Similar to the ii^-means algorithm, nearest neighbours classification re-
quires a distance measure d(-, ) to be defined over the data space A'. Let k
be some non-zero integer; then the /i:-nearest neighbours algorithm classifies
a new datum x (not in the training set X) as follows:

Algorithm 2.8: A:-Nearest Neighbours Classification Algorithm

1. Compute the distance d(x,Xi) between x and every element in the training set Xi,
2 = 1 , . . . ,?n.
2. Select the k training data Xj such that the distances d ( x , X i ) are the smallest (they
are called the A:-nearest neighbours of x ) .
3. Assign to x the most represented class label among the fc-nearest neighbours.

An important example of Algorithm 2.8 is when y {1; 1} (two-class

case), where one generally chooses A: as an odd number so as to ensure that
one class label is more represented than the other. This case is illustrated
in Fig. 2.15 with k = 5. Further insight into this method can be found
in [161].

D
D
D
D

Fig. 2.15. /c-nearest neighbours classification of two-dimensional data in the two-

class case, with k = 5. The new datum x is represented by a non-filled circle.
Elements of the training set (X, Y) are represented with dots (those with label 1)
and squares (those with label -hi). The arrow lengths represent the Euclidean dis-
tance between x and its 5 nearest neighbours. Three of them are squares, which
makes x have the label y = -hi.
 2 An Introduction to Signal Processing 61

2.4.7 Linear Discriminant Analysis

Linear Discriminant Analysis (LDA) is a supervised classification approach,
also called Fisher linear discriminant In its simplest form, it is devoted to
the two-class case where 3^ = {1; 1}. This classification technique has some
connections with principal component analysis (see Section 2.4.2, p. 54) in the
sense that it reduces the dimension of the data. In two-class linear discriminant
analysis, the dimensionality of the data is reduced to one: the direction of the
space that best separates the classes. More formally, we want to transform
the dataset (X, Y) = {(xi, y i ) , . . . , {yLm^Vm)} into a one-dimensional dataset
by setting
C^ = w"^Xi, z = l,...,m. (2.100)
This consists of projecting the data onto a fine with direction given by w.
In order for w to be effectively the best projection line for class separation,
we want to make sure that it maximizes the Fisher ratio (also known as the
Rayleigh quotient)

where U^^^^ (respectively ^ x ^ ^ - i ) ^^ defined as in (2.86), by only consid-

ering the training data with label y = + 1 (respectively y = 1), and similarly
for Mx^^+i ^nd A*x^=_i- The Fisher ratio is computed as the squared distance
between the means of the projected data d (^ = 1 , . . . ,m) divided by their
variances, and can be understood as a scaled distance between the training
data classes. The vector that minimizes J(w) can be computed analytically
as

WLO. = {s^,Ui+^xr/=-i)"'Ui - Mxr;=-i)- (2-102)

In practice, two-class LDA classification is implemented as in Algorithm 2.9
below. See [161] for additional details. In multi-class classification, LDA can
be also used to reduce data to C 1 dimensions, where C is the number of
classes.

Algorithm 2.9: Linear Discriminant Analysis

1. Compute the empirical means Mx'r=+i ^"^ M x 7 = - i ^^^^ ^^^ training data, as in
(2.86).
2. Compute the empirical covariance matrics U^J!^_^-^ and ^x'|y=-i ^^^^ ^^^ training
data, as in (2.86).
3. Compute the optimal projection direction WLDA according to (2.102).
4. Compute the projected data from the training set by d = WLDA^I , i = 1 , . . . , m
5. Implement any one-dimensional classification technique in order to classify the
projection C = WJ'DAX of a new datum x by using the mapped training set
{(Ci,yi),.--,(Cm,ym)}.
62 Manuel Davy

2.4.8 Bayesian Supervised Clsissification Using Generative Models

In this subsection, we assume the set of data X = { x i , . . . , x ^ } , as well as

unobserved data, to be classified following a known probabilistic model. In the
context of classification, such models are said to be generative because they can
be used to generate data by changing the parameters and the noise. Assume
for example that data are sinusoidal signals in noise; see (2.31). Signals with
label y = 1 have frequency A:(i), and signals with label y == 1 have frequency
A:(_i) with A;(_i) ^ A;(i). These models can be turned into likelihood functions
of the form (2.32) for each class. More generally, to each class with label y y
corresponds a likelihood function p(x|^y,y), denoted p(x|0y) for brevity.
Standard Bayesian classification consists of first learning the class pa-
rameters Oy for each class label y using (2.103), and then computing the
class likelihood of a new datum x using (2.104). In other words, given class-
priors p(0y|y), the posterior p(^y|Xy) is computed using class-training sets
Xy = {xi X such that y^ = y}, as follows:

p{dy\Xy) <x p(0y|y) J ] pi^il^v)' (2-103)

XiGXy

In the sine signals example, the training set is cut into X(i) and X(^_i)
containing respectively the signals with frequency A:(i) and k(^_iy The posterior
of A:(i) (respectively A:(_i)) is computed independently for each class.
Whenever the posterior is computed, it is possible to implement the clas-
sification of X via the computation of the class likelihood

p(x|Xy) = /p(x|0y,y)p(0y|Xy)d0y. (2.104)

Je
The integral (2.104) can be computed in analytic closed form in some cases;
in all other cases, MCMC enable its computation. Finally, assuming the 0-1
loss function, classification is performed by assigning x to the class for which
p(x|Xy) is maximum.
This technique requires some refinements for practical implementation be-
cause the parameter to learn may vary from one datum to another inside a
given class. In this case, it is not possible to learn the value of the parameter
of each class; we need to learn instead its density function. In this framework,
Oy becomes a set of pdf parameters, which is now denoted i/jy to distinguish it
from Oy used in the likelihood function, t/? is called the class hyperparameter
with prior p(t/?y|y). Consider again the sine signals example and, e.g., the class
y = 1. This class might include signals with frequencies close to ki. Assum-
ing known amplitudes and noise variance, the signal unknown parameter is
Oy = ki. Choosing a Gaussian distribution A/'(0y;/Xy, 27y) for the parameter
Oy in each class, the hyperparameter i/^y is composed of fiy and Uy (see [121]
for further details).
 2 An Introduction to Signal Processing 63

2.4.9 Hidden Markov Models

Hidden Markov models (HMMs) are widely used in speech recognition mainly
because they offer a robust pattern recognition dynamic scheme. HMMs in-
clude continuous state space systems such as (2.79)-(2.80): The parameter
vector On hves in a continuous state space 0; its evolution follows a so-called
Markov process (in the sense that the pdf of On only depends on the pdf
of On-i at each time n = 1,2,...); and the state cannot be observed di-
rectly (it is hidden). In the speech and audio processing literature, HMMs
mostly refer to finite, discrete state space dynamic models', that is, On is a dis-
crete random variable in a finite space with E possible state values, namely
0 = { e i , . . . , CE}', see Fig. 2.11. A finite, discrete state space HMM is governed
by
the state transition probabilities P{On = ei\On-i = ej) for (i,j) =
1 , . . . , Ethis is the discrete equivalent of (2.79);
the state likelihoods p(xn|0n = e^), i = 1 , . . . , Eequivalent to (2.80);
and the initial probabilities P{Oi).
Similar to the particle filtering problem, the issue is here to estimate the
sequence of states over time. Typically, in speech recognition the observations
are MFCCs extracted from frames over the speech signal, and the state is
the phoneme pronounced by the speaker. Likelihood functions are typically
Gaussian mixtures over the MFCCs (there is one GMM for each possible
state). Transition probabilities and GMM parameters are learned from a large
database and from the speaker's voice.^^

The Viterbi Algorithm

The aim of the Viterbi algorithm is the estimation of the sequence of states
Oi:T from time 1 to time T by maximum a posteriori. In other words, given
a sequence of observations Xi:7^, the Viterbi algorithm finds the sequence of
states Oi:T such that

di:T = argmax P(^I:T|XI:T), (2.105)

Oi.T

where the posterior probability of a sequence of state Oi^ is

T
P(01:T|X1:T) OC P ( 0 I ) H P(Xn|0n)P(0n|0n-l). (2.106)
n=2
In this context, estimation of the state is also called sequence decoding. The
Viterbi algorithm is presented below.
^^ Additional details may be found in [534], which proposes a tutorial on HMMs
and describes several basic algorithms such as the Baum-Welch algorithm for para-
meter learning.
64 Manuel Davy

Algorithm 2.10: Viterbi Algorithm

1. Initialization:
for i = 1,...,E, set wi{ei) = p(xi|^i = ei)P{Oi ~ e^).
2. Iterations: Forn = 2 , . . . , T ,
For z = 1 , . . . , ", compute

Wn{ei) = p ( X n | ^ n == Ci) jnax 'Wn-\{ej)9{Bn = ei|^n-i = e^) . (2.107)

For z = 1,... ,E, set '\\)n{ei) argmax u;n-i(ej)P(0n = ei|^n-i = e^).

e j j = l,...,;

3. Termination:
Compute OT = max i/;T(et).
4. State sequence backtracking

For n = T 1 , . . . , 2,1, extract the estimate at time n by 0n = ^ n ( ^ n + i ) .

The Viterbi algorithm implements an exhaustive search method along all

possible paths. Its computational complexity is 0{E^T), which makes it a
very efficient decoding algorithm, where the notation 0{E^T) means that the
computation time is proportional to T and to the square of E when other
factors are kept fixed. Note that Algorithm 2.10 has some similarities with
particle filtering, in particular in the way the weight Wn{^i) is computed;
compare (2.107) with (2.81).
Sparse Adaptive Representations for Musical
Signals

Laurent Daudet^ and Bruno Torresani^

^ Laboratoire d'Acoustique Musicale, 11 rue de Lourmel 75015 Paris, Prance

daudetQlam.jussieu.fr
^ Laboratoire d'Analyse, Topologie et Probabilites, CMI, Universite de Provence,
39 rue F. Joliot-Curie, 13453 Marseille cedex 13, Prance
Bruno.TorresaniQcmi.univ-mrs.fr

3.1 Introduction

Musical signals are, strictly speaking, acoustic signals where some aesthet-
ically relevant information is conveyed through propagating pressure waves.
Although the human auditory system exhibits a remarkable ability to interpret
and understand these sound waves, these types of signals cannot be processed
as such by computers. Obviously, the signals have to be converted into digital
form, and this first implies sampling and quantization. In time-domain dig-
ital formats, such as the Pulse Code Modulation (PCM)or newer formats
such as one-bit oversampled bitstreams used in the Super Audio CDaudio
signals can be stored, edited, and played back. However, many current signal
processing techniques aim at extracting some musically relevant high-level in-
formation in (optimally) an unsupervised manner, and most of these are not
directly applicable in the above-mentioned time domain. Among such seman-
tic analysis tasks, let us mention segmentation, where ones wants to break
down a complex sound into coherent sound objects; classification, where one
wants to relate these sound objects to putative sound sources; and transcrip-
tion, where one wants to retrieve the individual notes and their timings from
the audio signals. For such algorithms, it is often desirable to transform the
time-domain signals into other, better suited representations. Indeed, accord-
ing to the Merrian-Webster dictionary,^ to 'represent' primarily means 'to
bring clearly before the mind'.
Among alternate representations, the most popular is undoubtedly the
time-frequency representation, whose visual counterpart (usually without
phase information) is the widely used 'spectrogram' (see Chapter 2). Here,
at least visually, higher-level features such as note onset, fundamental fre-
quency or formants can be distinguished and estimated. However, a major

^See http: www. m-w. com.

66 Laurent Daudet and Bruno Torresani

consideration for time-frequency representation is that we have to decide on

a size for the analysis window, and it can be seen (Fig. 3.1) that this choice
can allow radically different features to be emphasized. Among all possible
choices for the analysis window, it is generally not possible to find the one
that is simultaneously optimal for all the features we want to extract.

0.05 0.1 0.15 0.2 0.25 0.3

Pure frequency domain

Time-frequency domain

Long analysis window

Time-frequency domain

Sfiort analysis window

Pure time domain

Fig. 3.1. Although equivalent, these four representations of the same signal (impulse
response of an open tube, kindly provided by M. Castellengo) highlight different
features of the signal (for the sake of clarity, only the magnitude of complex values
has been displayed).

The aim of this chapter is to introduce more advanced techniques, where we

put in strong assumptions about our signals: better performance is obtained,
but of course this is at the cost of a loss of generality. However, in many
applications mentioned above, there is a model for our class of analysed signals.
 3 Sparse Adaptive Representations 67

Within this framework, the representation of a sound is simply the set of

parameters for the model. For a given information extraction task, 'good'
representations are the ones where the parameters are simply related to the
desired high-level feature.
Here, representation methods can be divided into two large classes:
Analysis/synthesis methods: from the set of parameters (or coefRcients),
it is possible to reconstruct the original time-domain signal, either in an
exact manner or approximately.
Pure analysis methods: by focusing on only one type of feature that we
want to estimate, we allow some (irrelevant in this context) information
to be lost, and sound reconstruction is not possible anymore. For instance,
estimation of the fundamental frequency FO can usually be performed
without phase information.
In this chapter, we will focus on the former, in a typical analysis-by-
synthesis framework. The next section describes parametric representations
of signals, from the original McAulay and Quatieri sinusoidal modelling by
additive synthesis, towards most recent advances that also take into account
models for transients and noise. These parametric approaches are said to
be 'deterministic', in the way that the parameters are extracted locally in
the signal through deterministic algorithms. Alternate approaches tentatively
embed these parametric expansions into statistical models, where each sound
is seen as a particular realization of a random process having the hyper-
parameters we want to estimate. Section 3.3 is devoted to so-called waveform
representations, where the signal is seen as a linear combination of elementary
waveforms, chosen within a well-defined dictionary. This dictionary is usually
structured; i.e., its elements can be derived from a few elementary waveforms
through a number of simple operations such as modulations, translations, and
dilations. From a practical viewpoint, they are often chosen in such a way that
fast algorithms are available in order to get the expansion of a given signal.
After describing the general framework for waveform expansions, we will focus
on so-called sparse representations, in which the representation is compact:
few large coefRcients accurately represent the signal, while the vast majority
of the coefficients are small (hence neglected). Different models for sparsity
will be discussed, as well as measures for the distortion caused by neglecting a
large number of (small) coefficients. Within this framework, it is also possible
to enforce sound models such as { tones -h transients + noise }. In hybrid
representations, we decompose a sound as a union of orthonormal bases, each
of them being well suited for a particular type of feature (for instance a basis
of lapped local cosines for the tonal part, a basis of dyadic wavelets for the
transients). In such models, sparsity can be further exploited by taking into
account structures of significant coefficients, locally in the parameter (time-
frequency/time-scale) space.
Throughout this chapter, we will treat in parallel the discrete time and
continuous time settings. However, we found it necessary to be more explicit in
68 Laurent Daudet and Bruno Torresani

a few specific situations, and use either the discrete or the continuous setting.
We use the following notation: we reserve the letter n to denote discrete time
variables (i.e., we denote discrete time signals as x(n), n G Z), and the letter
t for continuous time signals x{t), t eR.

3.2 Parametric Representations

In this section we discuss some of the most widely used models for parametric
representations of sounds. Most of these originate from the additive synthe-
sis models, originally developed by McAulay and Quatieri [449] for speech
analysis and synthesis.

3.2.1 Sinusoidal Models and Relatives

The sinusoidal additive synthesis model [449] approximates a sound by a sum

of M sinusoids. In a discrete time setting,
M
^ W = 5 Z ^ ^ W ^^^ i^rn{n)) + e(n), (3.1)
771=1

where the phase (pm of the m-th partial is sometimes written as a primitive
of a smooth time-dependent frequency fm{t)-
pn/ks
^m{n)= I 27Tfm{t)dt-^(frn{0),
Jo
ks being the sampling frequency and {n) representing the error of the model.
Here, it is implicitly assumed that the amplitude and frequency parameters
(<^m? fm) for each partial sinusoid evolve slowly over time, in such a way that
their values can be estimated frame by frame (a typical frame size is 23 ms,
representing 1024 samples at a 44.1kHz samphng rate).
The selection of the partials is made first by peak-picking the magni-
tude of the short-time Fourier transform of the signal. Chaining the obtained
peaks into partials (i.e., curves n > fm(ji) for all m in a joint time-frequency
domain) is a non-trivial task which has received significant attention since
the early contribution of [449], including the hidden Markov chain approach
of [139], or linear prediction [383]. Whatever the chosen approach and al-
gorithm for partial chaining, the underlying idea is to exploit the supposed
smoothness of the time-dependent frequencies fm {t), and chain the peaks that
are close to each other in the frequency domain. Note that, for resynthesis us-
ing (3.1), the requirement of phase continuity at frame boundaries implies
some non-trivial interpolation scheme for the partials frequencies /m(0 (^^^
instance a cubic spline interpolation).
 3 Sparse Adaptive Representations 69

Time Time
Fig. 3.2. Local spectrum peaks, chained into partials (left). Partials grouped into
locally harmonic structures (right): partials corresponding to two locally harmonic
sources are represented by full lines and dashed lines respectively.

The partials may in turn be used for various purposes, including harmonic
source separation and possibly transcription, as shown in [652]. There, dis-
tances between time-dependent amphtudes am{t) and frequencies fmit) are
proposed, together with a measure of harmonic concordance between partials.
A perceptual distance between partials involving these three distances is then
built, and numerically optimized for grouping partials into locally harmonic
sources. An elementary example of such grouping is presented in Fig. 3.2.
This model has been refined by Serra in the spectral model synthesis (SMS)
model [576], where the residual of a sinusoidal model is also taken into account
as a so-called stochastic part. This modification has made it possible to perform
high-quality processing of general musical signals [649] (see also [408]) that
have a more complex behavior than speech signals. AppUcations that make
use of SMS range from audio effects (morphing, time-scaling, etc.) and source
separation to sound analysis (transcription).
One of the main limitations of the SMS approach is the lack of an explicit
model for the residual part. This results in having to keep this residual in the
time domain, hence requiring a very large number of parameters. As a first-
order approximation, this residual can be modelled as filtered white noise, the
parameters of the filter being estimated by the energy in perceptual frequency
bands [219] or through classical autoregressive (AR) or autoregressive mov-
ing average (ARMA) methods [472] (see Chapter 2). This solution is used in
the Harmonics and Individual Lines plus Noise (HILN) coder [532]. With fur-
ther improvements such as grouping harmonically related components, HILN
achieves fair sound quality at bit rates as low as 6 kbit/s.
However, in all the above-described modelling of the residual, there is
always the implicit assumption that the parameters evolve slowly over the
analysis frame. Obviously, this does not always hold; for instance it does
not hold at the sharp note attacks of many percussive instruments. These
fast-varying features are in certain cases characterized by sudden bursts of
noise, and/or fast changes of the local spectral content. These components
70 Laurent Daudet and Bruno Torresani

are usually referred to in the literature as 'transients'. When further accuracy

is needed, such as for high-quality coding, or for audio effects such as time-
scaling, an explicit model for transient sounds has to be included. Recently,
many three-layer models of the type STN { sinusoids -h transients + noise }
have appeared. For instance, in [269], the residual is transformed in a wavelet
packet basis that approximates the critical bands of the human auditory sys-
tem. In [641], transients are defined as dual components of the sinusoids,
and therefore modelled by additive synthesis in the DCT domain. More re-
cent transients modelling schemes for coding [569] distinguish between short
bursts of energy and sudden changes of signal level.

3.2.2 Parametric Random Models

The deterministic parametric models briefly described above may also be seen
from a probabilistic point of view. The advantage of such an approach is
twofold. On the one hand, it provides a way to model the intrinsic variability
of the phenomena under study. On the other hand, it also allows one to use
the powerful estimation algorithms that have been developed in the context of
statistical estimation. The price paid is of course an increased complexity of
the models. Chapter 7 presents a thorough presentation of the latter models;
see in particular (7.9) for the explicit several-notes model.
Remark 1. The model in (7.9) actually expresses the signal as a hnear com-
bination of 'elementary waveforms' of the form (/>^(n) cos(27r/^(n)n) and
(/)i(n)sin(27r/m(n)n), with unknown (random) parameters. This model may
be written in matrix notation; see (7.10) page 210. This creates, in some
sense, the connection between parametric harmonic models and sparse wave-
form models, to be discussed in the next section.

A nice feature of the approach above is the fact that given estimates for
parameters of the model, the latter may be used for signal synthesis. When
synthesis is not necessary, it is no longer necessary to start from the signal
waveform, and it may be easier to start from other representations of the
signal, for example a short-time Fourier spectrum. This approach has been
taken, among others, in [644], [645], where a new stochastic model for musical
instruments local spectra was introduced. The model parameters are first
estimated on a training set, and may then be used for transcription. Since only
spectra (and not signals) are modelled in this approach, separation may only
be performed by an appropriate post-processing, e.g. local Wiener filtering.
We refrain from discussing this approach in more details here, and refer to
Vincent [644] for a thorough description.

3.3 Waveform Representations

By waveform representations^ we mean representations of musical signals
as linear combinations of elementary waveforms^ or time-frequency atoms^
 3 Sparse Adaptive Representations 71

generated using simple construction rules. These include translations (i.e.,

time shifts) and modulations (frequency shifts) of a unique waveform for
'Gabor-type' waveforms, translations and dilations (rescalings) for 'wavelet-
type' waveforms, or more general transformation rules. Depending on the
problem at hand, and on the chosen approach, identifying the signal decom-
position may be more or less simple. We describe below a number of possible
approaches of various levels of difficulty. Even though it is difficult to give a
sensible classification of these approaches, we attempt to class them according
to two criteria:
1. Rigid vs. adaptive: Rigid techniques are characterized by the fact that
the signals under consideration are expanded onto a fixed family of wave-
forms (basis or frame of the underlying signal spacethese terms will be
explained below). Conversely, adaptive techniques choose the waveform
family as a function of the analysed signal, generally using greedy search
algorithms.
2. Redundant vs. strict: In 'strict' expansion techniques, the goal is to min-
imize the redundancy between the coefficients of the expansion. Strict
families generally correspond to orthonormal bases of the signal space.
Being strict often imposes severe constraints on the waveforms, and it is
sometimes more suitable to turn to redundant families (frames, dictionar-
ies, etc.), for which such constraints no longer apply. Redundancy (also
called overcompleteness since we are in finite dimension, or oversampling
in engineering terminology) offers a way to 'customize' the waveforms to
be used for the signal expansion. The price paid for redundancy is an in-
creased number of parameters (which can be penalizing for compression
purposes), and an increase in the complexity both at the analysis and the
resynthesis stages.

3.3.1 Rigid Expansions: Waveform Bases and Frames

We first make the distinction between non-redundant waveform systems (i.e.,

bases) and redundant ones (frames, dictionaries), before moving to distinc-
tions of more 'algorithmic' nature (rigid vs. adaptive).
Waveform representations generally make use of linear algebra language
and techniques, which we briefly recall here for the sake of completeness.
The reader can find more detailed descriptions in many classical mathe-
matics textbooks. We refer to [642] for a reference in the signal processing
context.
Signals are assumed to belong to some underlying vector space W, called
the signal space. H is generally assumed to be an inner product space which
is equipped with an inner product (,) : x^y H -^ {^^y) ^ C The latter
defines the norm \\x\\ of the signal x G H by ||a:|| = y/{x~x). Commonly used
signal spaces include
72 Laurent Daudet and Bruno Torresani

The (infinite-dimensional) L^ spaces of square-integrable functions;

L^{I) = | x : / -^ C, I |x(t)p dt < o o j ,

(/ being either the real line R or some subset), with inner product

{x,y) = j x{t)y{ty dt,

where '*' denotes complex conjugation;

The (infinite-dimensional) ^ spaces of square-summable sequences

e{Z) - J X : Z ^ C, ^ |x(n)p < oc I ,

I, n = oo )

with inner product

oo

{x,y) = Y^ x{n)y{ny ;
n = oo
The (finite-dimensional) space C ^ , with inner product
N-l
{x,y) = Y^x{n)y{ny .
n=0

Notice that in the latter case, signals are finite-dimensional vectors,

whereas in the first two cases, signals are infinite sequences or functions, which
may nevertheless be seen as 'vectors with infinitely many components'.

Waveform Bases

The simplest example of elementary waveform representations is provided by

orthonormal bases. There exists already a considerable amount of literature
dedicated to wavelet and local cosine bases (see e.g. [89], [113], [429], [642],
[667]), and therefore we do not go into details on the associated theory.
An orthonormal basis of an inner product space W is a family B = {(pi^ i ^
X} of elements of H ( J being an index set whose size equals the dimension of
the space W), such that
The family is complete in H; i.e., any signal x e H admits an expansion
of the form
E
x{t) = }_^aiip^{t) , (3.2)

For all i,i' e X,

{(Pi,(fi') = Sii' , (3.3)
where Su' is Kronecker's symbol, equal to 1 if i = i' and 0 otherwise.
 3 Sparse Adaptive Representations 73

Under such assumptions, it may be proved that the expansion (3.2) is

unique, and that the coefficients a^ in the expansion are nothing but the
inner products
ai = {x,^i), Iel. (3.4)
Finally, the latter satisfy Parseval's formula (this is an 'energy preservation'
formula)
^ K | 2 = ||xf. (3.5)

As mentioned above, the signal space may be finite dimensional (for exam-
ple finite-length discrete signals or finite-length band limited continuous time
signals), or infinite dimensional (for example finite or infinite support discrete
or continuous time signals), but the general framework remains the same. We
shall limit ourselves to spaces of finite-energy signals, i.e., L^ spaces in the
continuous-time case, and spaces of square-summable sequences in the discrete
and finite cases, with norms and inner products as above.
Remark 2. The reader who is not interested in the mathematical details may
simply remember that in the case of discrete time, finite-length signals (i.e., a
finite-dimensional signal space), expanding a signal onto an orthonormal basis
is nothing but applying a unitary matrix to the signal vector (a simple example
is the DFT matrix). The inverse operation, i.e., reconstructing a signal from
the coefficients of a basis expansion, is also a matrix-vector multiplication,
using the Hermitian conjugate (i.e., the complex conjugate of transpose) of
the transform matrix.
Among the orthonormal bases, wavelet and local cosine (with extensions
such as the modified discrete cosine transform, or MDCT) bases have been
particularly popular. The main difference between these two systems is the fact
that the time and frequency resolution of local cosine waveforms is uniform
in the time and the frequency axis, while wavelets offer finer time resolution
(and thus broader frequency resolution) at high frequencies.
Wavelets
Wavelets (see for example [113] and [429] for thorough reviews) are generated
from a single 'atom' i/; by regular translations and dilations. In the continuous-
time settings, a deep result by Mallat and Meyer states that it is possible to
construct (continuous-time) functions -0 G I/^(R) such that, introducing the
corresponding translations and dilations %l)jn of il) defined by
xjj^^it) = 2-^/2 IIJ{2-H - n) , t G M,
the family {ipjn, j,n G Z} is an orthonormal basis of L^(R). Therefore, any
signal X G L^(R) may be expanded in a unique way as

^(^)= m J2 i^^'^Jn)^Jn{i)-
-oo n= oo
74 Laurent Daudet and Bruno Torresani

i =2

i =4
i =5
Fig. 3.3. Dyadic tree of scale-time indices for a wavelet basis ^jn-

Here, j is a scale parameter, and n is a time parameter. The scale-time

index {j,n) is usually represented by a dyadic grid (see Fig. 3.3), which il-
lustrates the fact that the time sampling provided by wavelets depends upon
the value of the scale parameter j : at large scales (bottom of the tree), the
wavelets have larger support, and are sampled more coarsely than at small
scales (top of the tree).
Remarkably enough, such wavelet bases may (in most situations) be
generated through a mathematical algorithm called multiresolution analy-
sis (MRA), which is itself intrinsically connected to the sub-band coding
techniques of signal and image analysis. More precisely, a wavelet ortho-
normal basis and thus an MRA involve an auxiliary function 0 G I/^(M),
called the scaling function, such that the coefficients {x,(j)jn) of x G L^(R)
represent samples of a corresponding low-pass approximation of x at scale
2^ (the functions (/)jn are defined similarly to the wavelets ipjn)- Given
some (arbitrary) reference scale jo, it is possible to show that the family
{0jo,n5 ^ ^ ^ } U {i/^j^n, j < jo,n eZ} is another orthonormal basis of L^(R).
Therefore, the above wavelet expansion of a signal x may also be replaced
with another expansion

oo jo 1 oo

n= oo j= oo n oo

the basis functions (f^j^k^k G Z collecting the information contained in the

(doubly labelled) family of wavelets ipjk.k G Z, j > JQ. Given a wavelet basis
(or an MRA), the computation of the coefficients {x^^/jjn) of the expansion
of a signal x G L^(M) on the corresponding wavelet basis may be performed
through a sub-band coding algorithm, i.e., recursive filtering of the scaling
function coefficients and wavelet coefficients from a fixed pair of digital filters
(details on sub-band filtering may be found e.g. in [642]). The reconstruction
of the signal from its wavelet coefficients (x, il^jn) also rests on sub-band cod-
ing. The case of discrete signals (the discrete wavelet transform, DWT for
short) is handled by identifying signal samples x{n) with the scaling func-
tion coefficients (x, ^jm) of some 'underlying continuous time' signal, at some
reference (finest) scale j i .
 3 Sparse Adaptive Representations 75

Local cosines

Local cosine bases arose from an attempt to construct orthonormal bases

that 'look like' Gabor functions, i.e., waveforms generated by regular trans-
lates and modulates of a single waveform (see (3.10) below). For simplicity we
describe the case of infinite, continuous-time signals first, and then address
discretization issues. Starting from a family of compactly supported, positive-
valued window functions w^, essentially supported in intervals [an,n+i] (up
to rapidly decaying tails) subject to some technical conditions (see for ex-
ample [429], [667] for a detailed presentation) it was shown by Coifman and
Meyer that the corresponding family of functions

Unk{t) = J Wn(t) cos(7r(A:H-l/2)-^^^ , n E Z, fc G Z+,

yttn+l^n V Cin-\-l(^nJ
(3.6)
forms an orthonormal basis of L^(R). Therefore, any x G L^(R) admits a
unique expansion of the form
CO CX)

n= oo k=0

Here, n is a time index and /c is a frequency index. The usual practice^ is to

choose regularly spaced values of the a^, so that the local cosine atoms Unk
have constant length. In such cases, one may say that the local cosine atoms
are labelled by a regular grid in the time-frequency domain.
Local cosine bases adapted to continuous-time signals defined on finite in-
tervals are constructed similarly, modulo a mild 'twist' in (3.6) for treating
boundaries (essentially, the windows w become square windows at bound-
aries). Remarkably enough, a corresponding discrete theory has been devel-
oped, mainly based upon appropriate 'smoothing' of the windows used in
DCT-IV transform (a particular form of discrete cosine transform (DCT);
see [429] for example), yielding the modified discrete cosine transform (MDCT
for short). Again, Remark 2 above applies, and it is not necessary to master
the theory of cosine transforms and corresponding lapped transforms (see [431]
for a review) to use them practically.

Remark 3. Comparing wavelets and local cosines: As mentioned above, the

main difference between these two waveform systems lies in the fact that local
cosine atoms are constant size, variable shape functions, whereas wavelets
have constant shape and variable size. This is illustrated in Fig. 3.4, where
different wavelets and local cosine atoms have been represented, with various
sizes, frequencies, and locations. The two plots on the left each represent
^Although there are exceptions; for example, some audio coders use MDCT
atoms with two different lengths: wide windows w for describing partials, and narrow
windows in 'transient' regions.
76 Laurent Daudet and Bruno Torresani
0.2|
0.4 1

0.2

0 1 ^\V
^1 1
-0.2

-0.4
1 -
-0.6
n Q
200 400 600 800 1000 ^'^0 200 400 600 800 1000

0.2r

-0.1

1000 1000

Fig. 3.4. Comparing wavelets and local cosines. Daubechies 6 (top left) and
Vaidyanathan (bottom left) wavelets for three different values of the scale-time in-
dex (i, n), and narrow (top right) and wide (bottom right) local cosine atoms for
two different values of the time-frequency index (n, k).

three wavelets with different scales and position, using Daubechies wavelets of
order 6 (top) and Vaidyanathan wavelets (bottom).^ This illustrates a general
property of wavelet bases: the smoothness of the wavelet increases with the
size of its support.
On the right are represented local cosine atoms corresponding to two dif-
ferent window sizes (narrow window on the top plot, wide window on the
bottom plot) for two different locations and frequencies. In both cases, the
window function is compactly supported, which results in a poor frequency
localization for its Fourier transform.

Remark 4- Non-orthonormal bases, multi-bases, . . . : Many variations around

the theme of orthonormal bases in signal spaces are possible and have been
considered. One of the most popular is an approach in which the orthogonality
assumption is relaxed, while the completeness and non-redundancy assump-
tions are retained. As is often the case, relaxing one assumption yields more
^We refer to [113], [667] for detailed discussions of the common choices of wavelet
bases.
 3 Sparse Adaptive Representations 77

freedom in the construction of the basis, and such non-orthonormal bases may
be constructed from 'nicer' time-frequency atoms (for example, the atoms can
be made smoother, or more symmetric, or better time locaUzed). Another vari-
ant is obtained by assuming that the waveform basis is obtained from two or
more time-frequency atoms rather than a single one. Again, one obtains in
such situations 'nicer' time-frequency atoms. Examples of such a strategy are
provided by the multiwavelets, or the more general multiple bases (see for
example [17], [545]).

Frames and Redundant Representations

A major shortcoming of the (orthonormal or not) bases for application to

audio signal representation is the fact that orthonormality is often too strong
a constraint, in the sense that the corresponding waveforms do not have the
nice shapes one would like to see. They often have either poor time localiza-
tion or poor regularity (i.e., poor frequency localization). Even when some
assumptions are relaxed (for example, requiring biorthogonality rather than
orthogonality, or using multiwaveletsi.e., wavelet-like bases generated from
more than one basis function), the corresponding waveforms do not look nat-
ural. Also, while bases are clearly adapted for signal compression, representing
a signal by its coefficients with respect to a basis generally breaks some de-
sirable properties such as time-frequency covariance (see Chapter 2), which
is often useful for other tasks, including all signal analysis problems. Frames^
provide a way of introducing waveforms that are more 'natural', by introduc-
ing extra redundancy in the waveform system. They may also offer a way to
preserve the above-mentioned 'nice' properties, in particular time-frequency
covariance.
By definition, a family of waveforms {(^i,i T} is a frame of the signal
space H if the following weaker form of Parseval's formula holds^: There exist
positive constants ^ < A< B < oo such that for all x G 7i,

A\\xf<Y,\{x,'P^)?<B\\xf. (3.7)
iex

This imphes in particular that the frame is complete in H (as defined above),
and therefore that any x eH may be expanded as in (3.2). The main differ-
ence is the fact that the frame is generally not exacts which means that the
waveforms {^i^i G T} form a redundant system in H (one sometimes speaks of

^Notice that contrary to a common usage in the signal processing literature, the
term 'frame' does not represent a time interval, but rather a family of vectors in
a vector space. Since this terminology is also standard in mathematics, we shall
nevertheless use it here.
^While Parseval's formula (3.5) expresses energy conservation, such a weak Par-
seval's formula expresses energy equivalence, i.e., the fact that the energy of the
sequence of coefficients is controlled by the energy of the signal.
78 Laurent Daudet and Bruno Torresani

an overcomplete system). In such situations, the expansion (3.2) is not unique,

as there exists an infinity of choices for the set of coefficients {a^, i G X} yield-
ing signal expansions as in (3.2). The bad news is that the inner products
(x, (fi) can't be used as ai coefficients. Nevertheless, a wide variety of choices
is obtained by using a so-called dual frame., which is another family of wave-
forms {(^i,2 G 1} such that any x e H admits the following expansion with
respect to the ^i waveforms:

x{t) = Y,{^.^i)^iit)- (3.8)

iei
Notice that in such situations, one also has

iei
which may be interpreted as an inversion of the transform
X -^ {(x,(^i),i G X}. In both situations, the coefficients {x,ipi) or {x,(pi)
provide an alternate representation of the signal that often proves useful for
several signal analysis tasks.
The most classical choices for waveform frames are provided by the so-
called Gabor frames and wavelet frames^ whose basic theory is discussed in
detail in [113] (see also [184] and [261]). We shall hmit our discussion here to
Gabor frames and generalizations. The reader more interested in wavelet and
multiresolution frames is invited to refer to the vast literature on the subject
(see for example [89], [113] for tutorials).
Gabor frames, which actually correspond to sampled versions of the short-
time Fourier transform described in Chapter 2, have been fairly popular in the
musical signal representation community. Gabor frames are generated from a
unique window function g by time and frequency translations. In the discrete
time '^{Z) setting, the corresponding discrete Gabor functions gnk (where n
and k control time and frequency respectively) read
9nkil) = e2,.fcfe.(;-n3)^(; _ ^^^) (3 10)

Here, rig and ks represent time and frequency sampling rates, respectively. It
may be shown that under some mild assumptions, for any window g G ^^(Z),
the family {gnk^ n. A: G Z} is a frame of ^{Z) as soon as the product nsA:s
is small enough. The latter essentially controls the redundancy of the Gabor
frame: the smaller ng/cg, the closer the atoms. For example, increasing rig
reduces time redundancy, which may be compensated for by decreasing kg.
Many examples of signal processing applications of Gabor frames may be
found in the literature; see e.g. [185].
Gabor frames are often well adapted to musical signal processing, as they
provide 'direct access' to time and frequency variables simultaneously. Also,
using Gabor frames rather than the corresponding orthonormal bases^ allows
^A 'no-go theorem' known as Balian-Low phenomenon states that there cannot
exist 'nice' orthonormal bases of Gabor functions [113].
 3 Sparse Adaptive Representations 79

one to use 'nicer' (i.e., smoother and better localized) windows, and keep
translation invariance, which is a very important feature. On the other hand,
Gabor frames are made out of fixed resolution waveforms, which can therefore
not be adapted to the features of the signal. Hence, it is difficult to find in
such schemes a representation that would be well adapted to both transients
and partials of sound signals.
A good illustration of this fact is provided by the two images in Fig. 3.5
below, in which are displayed the Gabor frame expansions of a short piece
of guitar signal using two different windows. It clearly appears that a Gabor
frame generated using a narrow window (left image) is able to capture accu-
rately the transient parts of the signal, while a Gabor frame generated using
a wide window is much more precise for capturing partials. Neither of them
is able to do a good job for both types of components.

Time (s) Time (s)

10

^ 8

CP
^ A\

0.2 0.4 0.6 0.8 1.2 1.4 0.2 0.4 0.6 0.8
Fig. 3.5. Grey level images of two different Gabor frame representations of a short
piece of guitar sound. Left: narrow Hanning window. Right: wide Manning window.

From the above example, the following question arises naturally: Is there
a way of decomposing a signal into 'layers' that could be adequately repre-
sented by an appropriate waveform system (Gabor frame or other)? A very
interesting outcome of the frame theory is the fact that given a pair (or a
larger, finite family) of frames in a given signal space, their union is still a
frame of the same space, and thus suitable for expanding signals. Frames gen-
erated as unions of Gabor frames are called multiple Gabor frames. Multiple
Gabor frames have been considered in the literature for various purposes,
including source separation in signals and images, or musical signal process-
ing; see for example [150], [312], [690]. Examples discussed in these references
were generally based on a family of Gabor frames with identical windows at
different time scales. The goal in such situations is to be able to represent
80 Laurent Daudet and Bruno Torresani

transient components of sound signals using narrow Gabor functions, partials

using broad Gabor functions, etc.in a few words, find the best adapted
waveform for each component of the signal. An example of such a strategy
may be found in Fig. 3.6, where we displayed such a decomposition performed
on a harmonic sum of sine waves to which were added spikes with random
amplitude at random locations. The decomposition of this synthetic signal
was performed using the 'Time-Frequency Jigsaw Puzzle' method [312], to be
described below, which only exploits sparsity arguments. Even though the re-
sult is not perfect, we can see from the middle and bottom plots that the two
significantly different components were succesfully separated. The main point
of this example is the fact that broad Gabor functions were able to select the
partials of the harmonic components, while the narrow ones estimated almost
perfectly the transient components. Other illustrations of this technique on
real sounds may be found in [312] and [311].

10^

i ^ i l l / ^ ^ jA< kf U^^
1000 2000 3000 4000

1000 2000 3000 4000

1
'
" 1

10

11 1,1 It i h''f 1
TjjTT"^^ r 1
0 r 'f^-p ij
1 '
-10 ' 1 1

1000 2000 3000 4000

Fig. 3.6. Multilayered decomposition of a synthetic signal, obtained using the Time-
Frequency Jigsaw Puzzle technique, described in Section 3.3.2 below. From top to
bottom: original signal, tonal layer, and transient layer.

Another illustration of a similar strategy (using Gabor and wavelet fam-

ilies) may also be found in Figs. 3.11 and 3.12 below, where the tonal and
transient components of a vibraphone signal were succesfully separated using
structured approximation techniques (to be described below).

Remark 5. The need for adaptivity: In some sense, such multiple Gabor frames
provide a way to retain the best of the Gabor and multiresolution worlds. How-
ever, achieving such a program turns out to be quite difficult for practical
 3 Sparse Adaptive Representations 81

purposes, mainly because one has to deal with the extra redundancy intro-
duced by the use of several frames together. Indeed, given R Gabor frames
idnk^ (n, /c) G J , r = 1 , . . . , i?}, there is an infinite number of ways to expand
any x W as

^(o = E E "S^i^^'w. (3-11)

and one is naturally led to ask for the 'best' one, according to some given
criterion. In such a way, we are naturally led to leave the world of 'rigid
expansion' techniques, and enter the world of 'adaptive expansions'. To seek
the 'best' expansion, possible criteria are
Optimize the sparsity of the expansion: find a 'minimal set' of non-zero
coefficients such that (3.11) holds, at least within some prescribed accu-
racy. One is then naturally led to consider subsets of the multiple frame
one started with. These subsets can be frames themselves (the so-called
'quilted frames' introduced in [150]), or may not even be complete families
if adaptive methods are considered.
Organize the expansion in such a way that 'components' of the signal (for
example, transients, partials) are represented by large coefficients with
respect to a given waveform system that suits it.
In an ideal world, these two requirements would be equivalent, and they are
in fact nearly so. Unfortunately, coming up with an actual algorithm able to
perform such a task is far from easy. This again brings us to the problem of
sparse representations, to be discussed in Section 3.3.2 below.
Remark 6. We have only described above the multiple Gabor frame construc-
tion. Similar developments may be done starting from wavelet frames, the
union of bases, or the union of Gabor and wavelet systems (bases or frames).
In such situations, one talks of 'hybrid systems'.

Dictionaries
By 'waveform dictionary', one generally means a family of waveforms which
is more redundant than a frame. By definition, a dictionary in some signal
space H is a complete family of elements of 7Y, that is, a family such that
any signal x H admits an expansion as a linear combination of elements
of the dictionary. In infinite-dimensional signal spaces, dictionaries may even
not be frames, as they may contain too many elements for the right-hand side
inequality of (3.7) to be satisfied. However, when it comes to practical situa-
tions, i.e., finite-dimensional signal spaces, the dictionaries which are generally
considered in the literature are also frames, so that the distinction between
'frame methods' and 'dictionary methods' refer to the techniques that are used
to find the expansion of signals with respect to such systems rather than the
intrinsic properties of these systems. Therefore, we shall address 'dictionary
techniques' in Section 3.3.2 below.
82 Laurent Daudet and Bruno Torresani

3.3.2 Adaptive Expansions: The Quest for the Sparse Grail

It is an 'experimental fact' that in the case of audio signals, the expansions

described above are generally sparse: any signal in the considered class may
be represented with good accuracy by a truncated expansion of the type (3.2),
in which only the largest coefficients (in absolute value) are retained:
J
x(t)^^a,^.(^,^(t), (3.12)
j=i

where the indices i i , . . . zj have been chosen so that the absolute values of the
corresponding coefficients ai. are sorted in decreasing order.

Remark 7. There exist different ways of evaluating the sparsity of an expan-

sion. Essentially, any subadditive functional of the coefficients (i.e., giving less
importance to large values) would do the job. Classical choices are provided
by entropies, namely the Shannon or Renyi entropies^

H = -Y^ |a,|2 log2(|a,|2); R0 = J^^ log2 ( E I"'!"" ) ' (3.13)

iex ^ Kiel J
assuming that the sequence of coefficients ai has been normalized so that
^iei 1^*1^ ~ ^' I^^^yi entropies may be interpreted as sparsity measures for
0 < ;9 < 1. Moreover, it may be shown that the Shannon entropy is in fact
equal to the limit of the Renyi entropy as /? ^ 1.
Unfortunately, these different criteria generally yield different results, as
noted for instance by Jaillet [311]. Nevertheless, they turn out to essentially
agree in the case of simple signals, which motivated Jaillet to use them locally
in the time-frequency domain.
The fact that a signal admits a sparse expansion with respect to a given
waveform system also manifests itself by the fact that the histogram of the
coefficients of the expansion is significantly peaked at the origin (meaning
that a large number of coefficients are close to zero) and heavy tailed (a
slower decay of large coefficients). An example is provided in Fig. 3.8, where
the empirical probability density function (pdf) (computed here by an appro-
priate smoothing of the histogram) of the various representation coefficients
(time samples, Fourier samples, wavelet and MDCT coefficients) of two sig-
nificantly different audio signals are displayed. The two signals, a polyphonic
organ signal which is quite 'tonal', and a Castanet signal which is extremely
transient, are shown in Fig. 3.7. The pdfs in Fig. 3.8 represent various de-
grees of sparsity. As may be expected, the frequency and MDCT represen-
tations are better suited for the organ signal: the corresponding pdfs have
^Notice that Renyi entropies Rp are essentially a logarithmic form of the ^
norms \\ct\\2p = (X)i l^^l^^) of the normalized coefficient sequences, which have
also been used as sparsity measures.
 3 Sparse Adaptive Representations 83

5 Time (s)

Fig. 3.7. Two sample musical signals: a castanet signal (top) and an organ signal
(bottom).

10" 10"

n In ' ' uln *" "'''''Ifl

10" 10

-20 20

rS
10"

10"
I
10"'
-20 20

Fig. 3.8. pdf of various representations of the two sample signals of Fig. 3.7: ln(P(Q;))
vs a. Castanet (solid line) and organ (dashed line): Top: time samples and Fourier
coefficients; Bottom: wavelet and MDCT coefficients.

a m a x i m u m at the centre which is more sharply peaked, meaning t h a t one

has a large number of very small values, and a smaller number of significant
values.
84 Laurent Daudet and Bruno Torresani

Exploiting the sparsity of a waveform expansion turns out to become

tricky in practice if one wants to do it carefully. First, in the simplest
case of a waveform basis expansion, the selection of the number J of ele-
mentary waveforms to retain in (3.12) is not a simple issue. If redundant
waveform systems are preferred, the non-uniqueness of the representation in-
deed introduces more flexibility, but also yields more difficult decision prob-
lems. We describe below a few approaches that have been proposed in the
literature.

Matching Pursuit and Orthogonal Matching Pursuit

The Matching pursuit approach [430] (MP) is in the class of so-called 'greedy'
algorithms, i.e., according to the Wikipedia Encyclopedia,^^ an algorithm
which follows the problem-solving meta-heuristic of making the locally opti-
mum choice at each stage with the hope of finding the global optimum. The
MP is an iterative procedure that aims at approximating a signal through a
weighted sum of atoms such as in (3.12), where the atoms belong to a given
redundant dictionary V.
The basic principle of MP is as follows:

Algorithm 3.1: Matching Pursuit Algorithm

1. Initialization: Compute all the inner products ax = {x,u\).

Let ro(t) = x{t) and i = {).
2. Find maximum absolute modulus amongst all inner products:
\i = argmaxA |aA|.
3. Update the residual by subtracting the corresponding atom:
ri+i{t) = ri{t) - ax^uxiit).
4. Update all the Inner products:
ax = {ri+i,ux).
5. If stopping criteria is satisfied, then stop; otherwise update
i <r i-^ I and iterate to step 2.

Depending on the application, many stopping criteria can be used, for

instance a condition on the total number I of iterations, or on the norm of
the residual ||ri-fi||. After / iterations, the signal is written as
/-I
x{t) = ^ax,ux,{t)-^ri^i{t), (3.14)
2=0

and approximated by the first term of the right-hand side of the above equa-
tion, provided the residual rj^i is small enough (in norm).

10
See http://en.wikipedia.org/wiki.
 3 Sparse Adaptive Representations 85

A few important remarks can be made:

In very redundant dictionaries, finding the maximum inner product can
be a lengthy operation, as can be updating of the inner products. Speed-
enhancing tricks can in some cases be implemented, such as storing the
cross inner products of all elements in the dictionary, or implementing a
suboptimal search through a hierarchical organization of the parameters.
Due to the orthogonality between the residual and the selected atom,
choosing the atom that maximizes the inner product is equivalent to choos-
ing the atom that minimizes the norm of the residual. A further refinement,
called orthogonal matching pursuit [119], guarantees that the approxima-
tion at any i-th iteration is the best (in the L^ sense) on the span Vi of
the i already selected atoms. This is done by recursively orthogonalizing
the family of the i previously selected atoms. It is important to keep in
mind that it is in no way guaranteed that this is the best z-terms approx-
imation, except under very stringent conditions on the dictionary and the
signal, i.e., when the signal is exactly made of a small number of atoms
from the dictionary, this case being of little relevance for most practical
audio applications.
The classical choices for dictionaries include the union of Gabor frames with
different window sizes, extensions of the latter with 'chirp' waveforms (chirps
are frequency-modulated waveforms), or the union of several bases such as
wavelet and MDCT bases. We shall come back to such choices later on, and
refer the reader to [255], [257] for more details.

Time-Frequency Jigsaw Puzzles

The time-frequency jigsaw puzzle (TFJP) algorithm [312] is another way

of obtaining 'good' multiple Gabor frame expansions (as described above)
through a greedy approach, but it is simpler than matching pursuit. The
idea is essentially to partition the time-frequency plane into time-frequency
'supertiles' i7ri, and to find the optimal Gabor frame within each supertile,
according to some sparsity criterion (such as the entropies described above).
Supertiles are rectangular subdomains in the time-frequency plane, whose
area is significantly larger than the time-frequency area of the windows (typi-
cally, 25 times larger). A supertile is said to be of size M x AT if it contains M
time sampling points of the wide window, and N frequency sampling points
of the narrow windows. A typical value for M and TV is about 5.
More precisely, start from a fixed family of Gabor frames (typically, two
Gabor frames, one with a narrow window function, the other one with a wide
window function). Given a supertile Qn of the time-frequency index set, the
corresponding Gabor frame coefficients with respect to all considered Gabor
frames are computed, normalized locally (so that their absolute moduli sum
up to unity), and corresponding entropies i?/3 (as in (3.13)) are computed.
The Gabor frame whose coefficient set is sparsest (within Qn) wins.
86 Laurent Daudet and Bruno Torresani

In such a way a subset of each frame is selected, and the signal may
be iteratively projected orthogonally onto the corresponding subspace of the
signal space (as in orthogonal matching pursuit). At each iteration, a residual
signal is produced, and processed in a similar way. The iteration stops when
the precision is considered satisfactory. So far, no known proof exists for the
convergence of this method. However, the convergence is quite fast in practice:
less than 20 iterations are needed to achieve 300 dB signal-to-noise ratio.
In addition, T F J P provides decompositions of signals into 'layers' as fol-
lows: Assuming that two frames G^^^ = {9s } ^^^ ^^^^ ~ idx / ^^^ consid-
ered, the algorithms provides an expansion of any signal x in the form

X:{t) = x^'\t) + x('\t), xW{t) = ^f3sgi'\t), x^'^t) = J2<^>^9^x\t) ,

seA xeA

where x^^^ (a:^^^) is the 'component' (termed 'layer') of the signal which has
been 'identified' by the frame G^^^ (G^^^), and A and A are (small) subsets of
the global index set (in general a subset of Z^).
The simplest instance of this method is based on a pair of two Gabor
frames, with significantly different window sizes; say, for audio signals, 5 ms
and 45 ms. The corresponding Gabor atoms, when used in the framework of
the T F J P method, identify nicely partials and transients. An illustration of
such a strategy on a simple synthetic signal may be found in Fig. 3.6 above.
As long as the signal can be correctly modelled as a superposition of partials
(with slowly varying amplitude and frequency) and transients, T F J P is able
to identify and separate them. In the presence of more complex phenomena,
the method should be refined, and should include different types of atoms (for
example chirps). However, as is well known from matching pursuit approaches,
enlarging the dictionary of atoms does not necessarily improve the accuracy
of the identification of signal components: the more redundant the dictionary,
the larger the ambiguity of the selection.
The basic principle of T F J P (see [312, variant 2]) is presented in Algo-
rithm 3.2 below.

Algorithm 3.2: Timefrequency Jigsaw Puzzle

Choose a value for P G (0,1]. Choose two windows g^^\ g^"^^ and corresponding
sampling lattices; choose supertiles. Define a maximal number of iterations / and a
precision threshold e.

1. Initialization: Set ro(t) x{t) and i = 0.

2. Main loop: While z < / and \\ri\\ > e,
Compute coefficients {ri^g^l) for the two windows j = 1,2. Compute entropies
Rp for both windows within each supertile.
Select supertiles for which window # 1 yields the smallest entropy. Reconstruct
corresponding contribution x[^^ to layer 1.
Setri^i/2(t) = ri(t)-x\^\t).
Compute coefficients {ri^i/2,gJk) ^^^ ^^^ ^^^ windows j 1,2. Compute
entropies Rp for both windows within each supertile.
 3 Sparse Adaptive Representations 87

Select supertiles for which window # 2 yields the smallest entropy. Reconstruct
corresponding contribution x] to layer 2.
Setn+i(t) = r,^i/2(t)-xf\t).
3. Reconstruct layers i and 2 by summing up contributions xi'^ and x^-*, respectively.

3.3.3 Unstructured and Structured Hybrid Representations

In the approaches described above, the coefficients of the expansion of the

signal with respect to the waveform system are treated individually, and only
their magnitude is taken into account. However, in practice, it rarely happens
that a signal is characterized by isolated (in the index set) large coefficients,
with respect to any known waveform system. It could be so if one could use,
for expanding the signal, dictionaries of waveforms that make sense as ele-
mentary sound objects, but such a goal seems quite far away for the moment
(see nevertheless [43]). Therefore, sensible elementary sounds are rather made
of several chained elementary time-frequency atoms, forming time-frequency
molecules. Estimating such time-frequency molecules is the goal of what we
call structured approximation. Notice that structured sets of waveforms al-
ready appeared in Section 3.2, when we discussed the random parametric
models.

Structured Bases and Frame Representations

Structured basis approximation techniques have already been proposed in

different contexts in the signal processing literature. A good and famous ex-
ample is provided by the so-called embedded zero-tree wavelet (EZW) al-
gorithm [582] for image coding, which exploits the binary tree structure of
orthonormal wavelet bases (see Fig. 3.3). This algorithm is based on the fol-
lowing 'experimental fact': When a wavelet coefficient corresponding to a node
of the coefficient tree is zero or very small, then the coefficients attached to
the corresponding subtree are likely to be zero as well, yielding 'zero trees' of
wavelet coefficients.

Fig. 3.9. Subtree (in black) of the dyadic tree of Fig. 3.3 (suppressed edges appear
as dotted lines).
88 Laurent Daudet and Bruno Torresani

Similarly, structured sets of MDCT functions have also been considered

in [114], [469] from an audio coding perspective, exploiting the fact that when
an audio signal is expanded onto a local cosine basis, it often happens that
large coefficients form 'time-persistent' sets, or ridges (in the same spirit as the
chained peak sets in the Fourier domain described in Section 3.2; see Fig. 3.2).
In fact, as a result of the poor frequency localization of MDCT atoms, such a
localization is not so sharp. It has been shown that sine waves rather manifest
themselves in the MDCT domain by 'tubes' of three (or fewer) consecutive
(in frequency) ridges of significant coefiicients (see Fig. 3.10).

Time
Fig. 3.10. MDCT coefficient domain, and corresponding tubes of significant MDCT
coefficients.

Such 'persistent structures' may be exploited in different ways, in the

framework of Hybrid waveform audio models. In such contexts, the signal
is sought in the form

^(0 = J2 ^^^^(^) + Y^ (^xMt) + r{t) , (3.15)

6eA xeA
where the atoms I/JX (respectively us) are wavelets (respectively MDCT
atoms), the set A (respectively A) is a small subset of the wavelet (respec-
tively MDCT) coefficient set, and r is a (hopefully small) residual. Notice that
this model is again of the type 'tonal -\- transient -f noise'.
Practically, given a signal x, the goal is to estimate the significance maps
A and A, estimate the corresponding coefficients, and reconstruct the corre-
sponding layers of the signal

M) (t) = Y, ocx^t). x^^\t) = ^ psusit) ,

AGA 5eA

and the residual signal.

In the simplest approaches, the significance maps are estimated by thresh-
olding the observed wavelet and local cosine coefficients

ai = {x,2pi) , bj = {x,Uj) ,

the laxter being processed further to get estimates for the coefficients ax
and f3s' This type of approach, which was taken in Berger et al. [34] and
 3 Sparse Adaptive Representations 89

Daudet et al. [115], does not yield structured approximations, the coefficients
being processed individually. In the above-mentioned references, an iterative
approach was chosen, in which the MDCT layer was estimated first and re-
moved from the signal prior to the estimation of the wavelet layer. The diffi-
culty is then to provide a prior estimate for the number of MDCT coefficients
to retain for estimating the tonal layer. To this end, a transientness index was
proposed, based on entropic measures [468]. The latter actually provides an
estimate for the proportion of wavelet versus MDCT coefficients present in
the signal. Using this ingredient, the (unstructured) hybrid model estimation
procedure presented in Algorithm 3.3 is obtained.

Algorithm 3.3: Hybrid Model Estimation Procedure

Fix an overall number of coefficients J to be retained. Choose a wavelet basis and a
MDCT basis.
1. Estimate the number JM of MDCT and Jw of wavelet coefficients to retain.
2. Compute the MDCT coefficients of the signal as bm = {x,Um)', select the
JM largest ones (in magnitude) bmi, -bmj and construct the tonal estimate

Xton (t) = ^ brrij Uruj (t)

3. Substract the tonal estimate from the signal to get the non-tonal estimate Xnton(^) =
X{t) -Xton(^).
4. Compute the wavelet coefficients of the non-tonal estimate an = (xnton, V^n); select
the Jw largest (in magnitude) ones a m i , , . a m j 3"<^ construct the transient
Jw
estimate Xtrans(0 = y^^an^Un^jt).
j=l
5. Substract the transient estimate from the signal to get the residual estimate
Xres\t) = 3:^nton(^) ~ ^^transV^j-

Berger et al. [34] also suggested a greedy approach in which several passes
of this two-step procedure are expected to yield more precise estimates for
the layers x^^^ and x^^^. More precisely, a first estimate of the tonal layer
is obtained by peaking the largest coefficients of an MDCT expansion. This
estimate is substracted from the signal, and a first estimate of the transient
layer is obtained from the largest wavelet coefficients of this residual. The tonal
estimate is then updated by peaking the largest coefficients of the MDCT
expansion of this 'second-order residual', and so on. The difficulty of such
approaches is mainly in answering the question 'how many large coefficients
should one keep at each step?' The transientness index alluded to above could
perhaps be used at this point, but to our knowledge this has not been done
up to now.
To estimate structured significance maps, coefficients have to be processed
jointly rather than individually. In [115], a functional on the space of connected
subtrees of the wavelet tree is proposed. Numerical optimization of this func-
tional yields estimates for significance trees A of wavelet coefficients, and thus
90 Laurent Daudet and Bruno Torresani

for the wavelet layer. Similarly, a functional on the space of time-frequency

'tubes' is exploited in [114], yielding estimates for the MDCT layer. The corre-
sponding algorithm is again a two-step algorithm: A structured tonal estimate
is first obtained from a MDCT expansion, and substracted from the signal.
A structured transient estimate is then obtained from the wavelet coefficients
of the non-tonal signal. Finally the residual is obtained by substracting the
transient estimate from the non-tonal signal.
The latter approach has also been modified and refined in [469] to incor-
porate a more specific structured waveform model in the expression (3.15).
The main idea is to introduce a stochastic model for the significance maps A
and A that implements the desired persistence properties: time persistence
for A and scale persistence for A.
Let us first study the tonal layer. For the corresponding tonal significance
map, a standard first-order Markov chain model is used: For a given value of
the frequency index k, the membership probability of the index (n, k) to A is
governed by persistence (conditional) probabilities

TTk = P[(n,fc) G A|(n - l,fc) G A], ^ = P[{n,k) ^ A | ( n - l,fc) ^ A] .

The corresponding observed MDCT coefficients of the signal bnk = (x.Unk)

are distributed following zero-mean normal distributions, with large variance
CTj^ when the considered index {n^k) belongs to A, and small variance cr^
when (n,/c) does not belong to A. Given the parameters of the model (i.e.,
the variances and the persistence probabilities), the likelihood may be com-
puted explicitly. Therefore, parameters of the model may be estimated using
some training material and maximum likelihood procedures (here, EM algo-
rithms perform quite well; see Chapter 2). Once the parameters have been
estimated, the significance maps can be estimated in turn, using dynamic
programming procedures or others (see [469]). The tonal layer of the signal is
then constructed as

6eA
and the non-tonal layer reads

Similarly, using the natural tree structure of wavelet coefficients (see

Fig. 3.9), the transient significance map (that is, the significance tree) is mod-
elled using a Markov tree, following the lines defined in [107]. In a similar
manner, the distribution of the significance tree is then governed by persis-
tence probabilities: at scale j , we denote by

TTj = P[{j - IJ) e A\{j,n) eA], e = 2 n , 2 n + 1,

the probability that a node of the tree belongs to the significance tree A,
assuming that its parent belongs to A. The corresponding observed wavelet
 3 Sparse Adaptive Representations 91

coefficients of the non-tonal signal (xnton,'0jn) are modelled as random vari-

ables, distributed according to a mixture of two Gaussian distributions: large
variance for the indices belonging to A, and small variance for the other in-
dices. Again, the parameters of the transient model may be estimated from
the observed wavelet coefficients in some training material, and the signifi-
cance trees may be estimated as well using the ML approach. From this, the
estimated transient layer is given by

xeA
and the residual reads

^resv^j ^^ *^nton(,^j ^transV^j-

The actual implementation of such an approach actually involves more com-

plicated decisions; for example, parameters feature (slow) time variations, and
therefore have to be estimated locally, within (large) time frames; the size of
these frames is one of the important parameters to choose, which is not always
an easy task (for example, for musical signals it clearly depends on the musi-
cal 'style'). We shall not go into details of these aspects, and refer the reader
to [469] for a more detailed account of the model, and [114] for an application
to audio coding.

Remark 8. Such two-step approaches have two major drawbacks:

As stressed before, the relative proportion of MDCT and wavelet coeffi-
cients to be picked has to be determined prior to the estimation of the
significance maps and trees. The entropy-based estimate proposed and
studied in [468] yields satisfactory results as long as the tonal and tran-
sient signal representations are sparse enough (i.e., the significance maps
A and A are small enough). However, a global estimation procedure would
be far preferable.
It is extremely difficult to obtain any error estimate in such schemes. For
example, the errors in the estimation of proportions of the two layers
should be taken into account, which does not appear to be easy.

An example of such multilayered decomposition is given in Fig. 3.11.

A vibraphone signal (about 6 seconds, sampled at 44.1kHz) has been decom-
posed following the lines of the Markov model described above (see [469] for
details). The original signal is displayed in the top plot, and the tonal and
transient layers are respectively represented in the middle and bottom plots.
The spectrograms of these three signals are shown in Fig. 3.12. As appears
clearly in the latter figure, the algorithm was able to capture the time per-
sistence (scale persistence) of the waveform coefficients representing the tonal
(transient) layer.
92 Laurent Daudet and Bruno Torresani

Transient
0.2
0 i' i ^ ^ \ ^
-0.2

0 1 2 3 4 5 6

Fig. 3.11. Multilayered decomposition of a vibraphone signal, obtained using the

Markov-based structured approximation techniques described in Section 3.3.3. fYom
top to bottom: original signal, tonal layer, and transient layer.

Structured Matching Pursuits

Within the MP framework, it is also possible to design 'structured' versions

of the basic algorithm presented above. The underlying principle is to select,
at every iteration, a whole group of atoms that have some relation in the
parameter space. In the harmonic matching pursuit (HMP) [256], this relation
is a harmonic frequency relation: ^^ the dictionary is made of a multiscale
family of Gabor atoms (compare with (3.10))

^s,n,fc(/) = e2^'^'^^(^-""^)^,(/-nns),

the window QS being a rescaled copy of ^, at scale s, and the indices {s, n, fc}
belonging to some fixed index set. A 'harmonic atom' is defined as a group of
M harmonically related atoms:

HGs,n,k{t) = {9s,n,k{^)}m=l. .M'

The search is restricted to an interval of fundamental frequencies ki < k <

fc2 such that all atoms forming a given harmonic atom can be considered to
be approximately orthogonal. This requirement is crucial since in that case
the signal energy carried by the harmonic atomwhich is now the criterion
^^The technique may easily be modified to account for any prescribed inharmonic-
ity, as soon as the relationship between partials is known.
 3 Sparse Adaptive Representations 93

100 200 300 400 500

Time

100 200 300 400 500

Time

100 200 300 400 500

Time

Fig. 3.12. Time-frequency representation (spectrogram) of the vibraphone signal

(top) and the two layers: tonal layer (middle) and transient layer (bottom).
94 Laurent Daudet and Bruno Torresani

for choosing the best atom at every iterationis (approximately) the sum of
the squared inner products of the individual atoms, i.e.,
M

The rest of the procedure is the same as classical matching pursuit.

In a similar way, the MP can be used to construct structured decompo-
sitions for a sines + transients -f- noise model, in the same spirit as before.
This is particularly adequate for sounds that are not harmonic and that have
sharp onset transients, such as percussive sounds. In the molecular matching
pursuit (MMP) [116], the dictionary is, as in the case of hybrid expansions
above, the union of a local cosine (MDCT) basis and a discrete dyadic wavelet
(DWT) basis. Here, one considers two kinds of structures, called 'molecules':
tonal molecules that are horizontal structures in the MDCT basis, and tran-
sient molecules that are local subtrees of the dyadic wavelet tree. At each
iteration, the algorithm looks for the strongest correlation among neighbour-
ing atoms, then identifies and subtracts the corresponding molecule. Again,
the procedure is iterated until some stopping criterion is met.
An example of molecular matching pursuit may be found in Fig. 3.13,
where the locations of selected atoms (wavelets and MDCT atoms) in the
time-frequency domain are shown. The signal is the vibraphone signal analysed
in Fig. 3.12, and MP and MMP were required to select the same number of
atoms. In Fig. 3.13, the first two panels (from top to bottom) correspond to
the standard MP, and a significant overlap between atoms may be observed.
The last two images correspond to the 'molecular' version, and the redun-
dancy between selected atoms has been greatly reduced. In addition, the time
(scale) persistence property in the MDCT (wavelet) domain appears much
more clearly in the bottom figures.

3.3.4 How Can Waveform Expansions Be Used for Transcription?

Each of the waveform representations described above can be used as a pre-

processing stage for the task of music transcription. However, transcription
is much more than just looking at an alternate representation of the signal:
proper transcription algorithms, as can be seen throughout this book, have
to include a lot of high-level heuristic rules, for instance in the frequency do-
main, for lifting the usual transcription ambiguities due to harmonic relations
between notes (octave 1:2, fifth 2:3, etc.) that have a large number of over-
lapping partials; or in the time domain, for making the distinction between
vibrato and the start of a new note.
A good representation of the signal is one that facilitates this note iden-
tification process, seen here as post-processing. In an ideal world, this would
be a rather straightforward task: when one projects a signal on a basis of
tonal atoms, the notes would simply be given by the large tonal atoms, and
 3 Sparse Adaptive Representations 95

8000

7000

6000

> 5000
c

. 4000

u- 3000

2000

1000

0
time (s)
7r

6 o o o o o o o o 0
5 o o o o o o o o o
15 4 o o o o oo no (O 0CD 0<XS>0 o
o
" 3o o o o o o OQEXX) c o o OQOO o
2 o o o o o oo GD OOOD 000 0 o
-o o o o O O 0

7000

6000
>^
^ 5000
Qi
2 , 4000
> - 3000

2000

1000

7
time (s)
6 O o o o o o <a o
5 o o o o o o OD o o
^
(d
4 o o o o o o o > o o
^ 3 o o o o o o o D o o
2 o o o o o o o a> o o
1
^ o o o o o o D o o

Fig. 3.13. Localization of selected coefficients in matching pursuit, in a standard

implementation (MP, top two plots), and in a structured 'Molecular' implementation
(MMP, bottom two plots), for the same total number of selected atoms. Prom top
to bottom: MDCT coefficients in MP, discrete wavelets in MP, MDCT coefficients
in MMP, discrete wavelets in MMP. The signal is the vibraphone excerpt analysed
in Fig. 3.12.
96 Laurent Daudet and Bruno Torresani

their significance maps would represent 'piano-roll' types of musical scores.

If one wants a precise description of the attack transients, they are given by
the large atoms on a transient-like basis (e.g., wavelets). However, practical
situations are far from being that simple (readers should not stop reading
this book after this chapter), mainly because atoms of a basis in general do
not look like musical notes: First, a basis of L^(IR) is not invariant through
time shifts and has a limited frequency resolution, and therefore cannot in
general accommodate a pure sine wave with only one atom per time frame.
This leads to the introduction of frames or redundant dictionaries of atoms,
therefore losing the uniqueness of the decomposition. Second, typical note du-
rations are much longer than the support of atoms, and on these long time
scales their characteristics vary due to energy decay, vibrato, and so forth;
therefore atoms have to be chained in some way into coherent structures
such as the molecules described in the previous subsection. Research is ac-
tively being carried out for the construction of 'nice' atoms that resemble
notes as closely as possible. However, it seems unlikely that significantly bet-
ter results could be obtained without being restricted to specific classes of
instruments.
Finally, the question that we will leave open is. How much of this empirical
information (harmonicity, time variations, rules for octave errors, etc.) has to
be put in the decomposition process itself, and how much is left for the post-
processing transcription? Until the last few years, most of the transcription
methods were based on simple decomposition schemes, such as the short time
Fourier transform, and the transcription effort was on the post-processing of
this data. The recent development of methods that provide sparse, overcom-
plete and/or structured decompositions now makes it possible to include a
lot of information in the decomposition process itself, with the hope that this
would provide some data that are easier to interpret in terms of musical ob-
jects. However, this is a very new field that still requires a lot of investigation.
In particular, by putting too much prior information on the sought-after ob-
jects there is the risk that we might enforce constraints that are too stringent,
and that do not reflect the variety of musical signals. A typical illustration is
given in Fig. 3.14, where we have extended the molecular matching pursuit in
order to look for harmonic groups of molecules, called meta-molecules [376].
On this very simple sound file, containing three notes played distinctly on a
clarinet, the algorithm has made the relevant grouping of partials, and the
notes can be correctly identified. However, for complex polyphonic mixes, the
algorithm fails to perform simultaneously a good tracking in time and a rel-
evant harmonic grouping. Clearly, this algorithm (which incidentally was not
designed for transcription) is going too far.
Finally, it is likely that future transcription systems will go beyond the
traditional one-way transform > post-processing^ and try to optimize the
system globally. For instance when a high-level notes hypothesis is made,
some further signal-adaptive analysis can be performed to (in)validate the
hypothesis.
 3 Sparse Adaptive Representations 97

1 - - MIM
^^^^^^^^^^^^^^^^^^
- ^ 4

IJHHaHHHHHHMHMHMHnX

4
g- 4000

[ -

'^" ^piwMMiawwwwMJi
1000 ^mmmmmmmmm J

1.2 1.3 1.5

Time [s]

Fig. 3.14. Result of the meta-molecular matching pursuit (MSP) algorithm on three
notes played by a clarinet. The molecules are shown as grey rectangles superimposed
on the spectrogram. Black boxes show the detected meta-molecules, which in this
very simple case correspond to the notes.

3.4 Conclusion

We have described in this chapter a number of approaches for musical sig-

nal representation, emphasizing parametric modelling and waveform expan-
sion approaches. Parametric models are intellectually appealing, as they may
sometimes be interpreted in terms of physical models for instruments, or more
generally, in terms of sound production. Waveform approaches do not allow
such easy interpretations, but offer other advantages, such as their flexibility
and their computational efficiency.
Interestingly enough, these two competing approaches tend to somewhat
converge. Indeed, as may be seen from Chapter 7 (see in particular (7.9)),
expanding the amplitude functions of sinusoidal models onto bases of spline
functions (for example) generate waveform systems close to Gabor systems.
Similarly, the introduction of 'structured approximation' techniques in the
waveform approaches may be understood as an attempt to move waveform
models in the direction of more physically realistic models, and recover the
possibility of identifying partials in the signal. In both cases, this convergence
98 Laurent Daudet and Bruno Torresani

may be interpreted as an attempt to keep the best of the two approaches,

while ehminating their shortcomings.
Another major separation between the waveform approaches we just re-
viewed is between rigid and adaptive approaches. While the first ones are
more computationally efficient, the latter offer more flexibility, as they exploit
redundant 'dictionaries' which generally contain more realistic waveforms. Un-
fortunately, flexibility has a price: sparse expansions in very redundant dictio-
naries are more difficult to identify. However, the introduction of structured
expansion techniques help to reduce the ambiguity, while being closer to para-
metric approaches.

Acknowledgements

This work was supported in part by the European Union's Human Potential
Programme, under contract HPRN-CT-2002-00285 (HASSIP), and the Math-
STIC program of the French Centre National de la Recherche Scientifique^
in the framework of the project 'Approximations parcimonieuses structurees
pour le traitement de signaux audio'.
 Part II

Rhythm and Timbre Analysis

Beat Tracking and Musical Metre Analysis

Stephen Hainsworth

Tillinghast-Towers Perrin, 71 High Holborn, London WCIV 6TH, UK

swh210cantab.net

4.1 Introduction

Imagine you are sitting in a bar and your favourite song is played on the
jukebox. It is quite possible that you might start tapping your foot in time
to the music. This is the essence of beat tracking and it is a quite automatic
and subconscious task for most humans. Unfortunately, the same is not true
for computers; replicating this process algorithmically has been an active area
of research for well over twenty years, with reasonable success achieved only
recently.
Before progressing further, it would be useful to define beat tracking
clearly. This involves estimating the possibly time-varying tempo and the
locations of each beat. In engineering terms, this is the frequency and phase
of a time-varying signal, the phase of which is zero at a beat location (i.e.,
where one would tap one's foot). When musical audio signals are used as an
input, the aim of 'beat-tracking' algorithms is to estimate a set of beat times
from this audio which would match those given by a trained human musician.
In the case where a notated score of the music exists, the musician is used as
a proxy for it (hopefully the musician's set of beats would align with those in
the score). Where no score exists, the musician's training must be accepted
to return a metre equivalent to how the music would be notated. Note that
this implies that it is the intended rather than the percieved beat structure
that is the focus here.
Beat tracking as just described is not the only task possible. Some algo-
rithms attempt only tempo analysisfinding the average tempo of the sam-
ple; others attempt to find the phase of the beat process and hence produce a
'tapping signal'. Meanwhile, some methods also attempt a full rhythmic tran-
scription and attempt to assign detected note onsets to musically relevant
locations in a temporally quantized representation. This is often considered
in terms of the score which a musician would be able to read in order to
recreate the musical example [352]. MIDI signals are also commonly used as
102 Stephen Hainsworth

inputs and, assuming that the signal is of an expressive performance, all of

the above tasks are again possible aims.
This chapter is organized as follows. Section 4.2 gives an overview of meth-
ods and approaches to beat tracking. However, as with any engineering system
which is trying to replicate a real-world process, it is useful to examine the
actual process before trying to build a model. Section 4.3 of this chapter
briefly discusses some of the musical background behind beat tracking. Next,
detection of onsets in musical audio signals is discussed in Section 4.4 be-
fore some of the more influential approaches to beat tracking are presented in
Sections 4.5 to 4.9. Probabilistic models are examined in more detail in Sec-
tion 4.10. Section 4.11 presents trials of various algorithms on a comprehensive
test database and conclusions will be drawn in Section 4.12.
There are many immediate and commercial applications of a successful
beat tracking program which have perhaps motivated some of the research.
Some of these are: automatic accompaniment of a solo performance [538],
synchronization of two music streams (e.g. for DJing [94]), correctly timed
recovery from CD skipping (see [660] for a similar application), intelligent
time stretching of musical samples [151], determination of good points for
looping algorithms (useful for studio samplers which are heavily utiHzed in the
creation of dance music) and adding tempo synchronous eff'ects. Other uses
include database retrieval [633] and metadata generation [566], provision of a
'rhythmic similarity' function to listeners (either in playback or for purchase
recommendation) and rhythmic expressiveness transformations (e.g. adding
swing to a musical example [244]). In addition, beat tracking can form a good
basis for any automated transcription program (e.g. [126], [231], [353], [611])
from which to begin its analysis.

4.2 S u m m a r y of Beat-Tracking Approaches

Beat tracking with computers has been an active area of research since the
early 1980s, though psychological models of human rhythmic perception pre-
date this. The early work was undertaken in the fields of music perception
and computer science, though the emphasis shifted towards engineering and
statistics as computing power increased.
As a result of this paradigm shift, the aims and approaches of the methods
described below vary considerably. It would hence be useful to categorize
them. The first and most important distinction is by type of input; most of
the earlier algorithms for beat tracking used a symbolic^ or MIDI input while
audio signals have been used more recently. This is at least partly because
the signal processing required to extract rhythmic cues from the audio was
beyond the power of early computers. It should be noted, however, that many
of the more recent methods implicitly convert an audio stream to a set of
MIDI-type inputs via the use of a pre-processing onset-detection algorithm.
^Symbolic data usually consists of a quantized set of note start times.
 4 Beat Tracking and Musical Metre Analysis 103

The second important differentiation between approaches is the intended

purpose of the algorithm. Much of the early work was conducted with the
music psychology goal of understanding how humans perceive music and at-
tempting to model this. Other approaches have goals based more in engineer-
ing and attempt to capture information in the signal without direct reference
to human perception. Specifically, those studies undertaken within the frame-
work of automated transcription attempt to return to the underlying score
rather than any human perception of the performance.
The next major distinction between the algorithms is the broad approach
used. Categorizations here could include
rule-based;
autocor relative;
oscillating filters;
histogramming;
multiple agent;
probabilistic;
though there are methods which do not fall neatly into any of these classes.
Descriptions of these six broad approaches can be found later in Sections 4.5
to 4.10.
Another, more subtle method of classifying algorithms is by causal [572]
operation. In a causal model, the estimate of the metre at a given time depends
only on past and present data. A non-causal model allows the use of future
data and backward decoding. Another way to consider it is that a causal
algorithm attempts to mimic human tapping and uses data only up to the
current time to decide whether a beat should be marked or not. Semi-causal
algorithms have also been produced where the estimate is made after a short
time-lag, typically around 20 ms. These can often give a 'strict' causal estimate
but at the cost of optimality.
Finally, the algorithms can be grouped by their intended output; some
only produce a best estimate of tempo while others evaluate phase as well,
therefore giving the beat. Gouyon [242] separates these into tempo induc-
tion, the estimation of the most likely tempo given a segment of data, and
beat tracking, which is the following of the beat through an extended ex-
ample. Some methods also extract the super-beat and/or sub-beat structure
(that is, slower and faster pulses than the beat, respectively), while some
only attempt estimation of either the super- or sub-beat and not the actual
beat.
Table 4.1 summarizes some methods found in the literature, indicating
the type of input used and any causal nature. Others which do not fall into
any particular category are Sethares and Staley [578], Smith [601], Miller
et al. [464], and Bilmes [37]. Two other studies which present surveys or re-
views of beat tracking are [243], [249].
104 Stephen Hainsworth

Table 4.1. Summary of beat-tracking methods. Key for Input column: A = audio,
M = MIDI, and S = symboUc.
Approach Author and year [Ref] Input Causal
1) rule-based Steedman 1977 [607] S
Longuet-Higgins & Lee 1982 [418] S
Povel & Essens 1985 [529] S
Parncutt 1994 [497] S
Temperley & Sleator 1999 [622] M
Eck 2000 [165] S
2) autocorrelative Brown 1993 [55] S
Tzanetakis et al. 2001 [632] A
Foote 2001 k Uchihashi [194] A
Mayor 2001 [445] A
Paulus k Klapuri 2002 [503] A
Alonso et al. 2003 [15] A
Davies & Plumbley 2004 [118] A X
3) oscillating filters Large 1994 [390] M X
McAuley 1995 [450] M X
Scheirer 1998 [564] A X
Toiviainen 1998 [626] M X
Eck 2001 [166] A
4) histogramming Gouyon et al. 2001 [245] A
Seppanen 2001 [573] A X
Wang & Vilermo 2001 [661] A
Uhle & Herre 2003 [635] A
Jensen & Andersen 2003 [318] A X
5) multiple agent Allen k Dannenberg 1990 [14] M
Rosenthal 1992 [546] M
Goto et al. 1994 [221] A X
Dixon 2001 [148] A/M
6) probabilistic Laroche 2001 [392] A
Cemgil et al. 2000 [75], [76] M
Raphael 2001 [537] A/M
Sethares et al. 2004 [577] A
Hainsworth k Macleod 2003 [266] A
Klapuri 2003 [349] A X
Lam k Godsill 2003 [386] A
Takeda et al. 2004 [617] M
Lang k de Freitas 2004 [387] A
 4 Beat Tracking and Musical Metre Analysis 105

4.3 Musical Background to Rhythmic Structure

Typically, music consists of sounds generated concurrently by a number of
different sources (usually musical instruments of varying kinds). These are
organized in a temporal manner, the structure of which forms the 'rhythm' of
the piece. Most music has a coherent temporal structure, as this is pleasing to
most listeners. Thus the rhythm of a piece more readily lends itself to analysis
than the harmonic structure, which can often be much more complex.
At the top level, the rhythm describes the timing relationships between
musical events within a piece. The Oxford English Dictionary [624] gives the
definition of rhythm as
a. The aspect of musical composition concerned with periodical accent and the
duration of notes.
b. A particular type of pattern formed by this.
Cooper and Meyer [102] define it as the way in which one or more unaccented
beats are grouped in relation to an accented one. The term metre is sometimes
used in an equivalent manner to rhythm, though in music psychology it takes
on a different meaning. Here, metre is the number of pulses between the more
or less regularly recurring accents in a piece of music [102]. Thus, the metre
is a constituent of the rhythm of a piece of music; however, the grouping of
accents into patterns and the interaction of this process and the metre are
closer to describing the rhythm of a piece.
Some further analysis can be made; Bilmes [37] breaks down musical timing
into four subdivisions. The first is the hierarchical metrical structure, which re-
lates the idealized timing relationships as they would exist in a musical score,
i.e., quantized to a grid.^ Next is tempo variation, which gives the possibly
time-varying speed at which the events are sounded. Another level of abstrac-
tion gives timing deviations, which are individual timing discrepancies around
the time-varying metrical grid (e.g. 'playing ahead of the beat'; swing^ can
also be considered a timing deviation). Finally there are arrhythmic sections,
where there is no established rhythm. These will be ignored from now on as
fundamentally impossible to analyse rhythmically, except as a collection of
unrelated note start times.
The metrical structure can also be broken down into a set of three hi-
erarchical levels. Klapuri [349] describes the beat or tactus as the preferred
(trained) human tapping tempo and is what most of the beat-tracking algo-
rithms attempt to extract at a minimum. This usually corresponds to the 1/4
note or crotchet when written out in common notation, though this is not
always the case: in fast jazz music, the pulse is often felt at half this rate (1/2
note or minim), while hymns are often notated with the beat given in minims.
^Dixon [148] uses the term 'scoretime', measured in beats since the start of the
sample to describe this representation.
^Swing is a style where the second l/8th note of every beat is slightly delayed;
it is a characteristic of jazz and some rock music.
106 Stephen Hainsworth

r > >r^
*5r ^ ^
tatum
beat

bar
timing

Fig. 4.1. Diagram of relationships between metrical levels.

However it is notated, the rate at which beats occur defines the tempo of the
music [404].
At a lower level than the beat is the tatum, which is defined to be the
shortest commonly occurring time interval. This is often defined by the l/8th
notes (quavers) or l/16th notes (semiquavers). Conversely, the main metrical
level above the beat is that of the bar or measure. This is related to the rate of
harmonic change within the piece, usually to a repeated pattern of emphasis
and also notational convention. Fig. 4.1 gives a diagrammatic representation
of the above discussion. Included is a set of expressive timings for the score
given. While obvious, it should also be noted that onsets do not necessarily
fall on beats and that beats do not necessarily have onsets associated with
them.
From here, metrical levels below the beat, including the tatum level, will
be termed the sub-beat structure, while the conversebar levels, etc.will be
labelled the super-beat structure. In between the tatum and beat, there may
be intermediary levels, usually related by multiples of two or three (compound
time divides the beat into three sub-beats, for instance). The same applies be-
tween the beat and bar levels. Gouyon [242] gives a comprehensive discussion
of the semantics behind the words used to describe rhythm, pointing out many
of the dualities and discrepancies of terminology. One point he raises is that
the terms beat or pulse are commonly used to describe both an individual
element in a series and the series as a whole.
An interesting point is raised by Honing [294], who discusses the duality
between tempo variations and timing: the crux of the problem is that a series of
expressively timed notes can be represented either as timing deviations around
a fixed tempo, as a rapidly varying tempo, or as any intermediate pairing. This
is a fundamental problem in rhythm perception and most algorithms arrive at
an answer which lies between the extremes by applying a degree of smoothing
to the processesthis usually means that estimated tempo change over an
analysis segment is constrained by the algorithm and any additional error in
expected timing of onsets is modelled as a timing deviation.
This leads to the concept of quantization, which is the process of assess-
ing with which score location an expressively timed onset should be associ-
ated. Here, score location refers to the timing position the onset would take
 4 Beat Tracking and Musical Metre Analysis 107

when notated upon a traditional Western musical score or other equivalent

representation. However, for most purposes, it can be reduced to the num-
ber of beats (and sub-beats) since the start of the sample. Quantization is
an important problem and other specific studies on this topic include Cemgil
et al. [75] and Desain and Honing [142].
The phase of the beat is determined by a series of stresses or accents,
termed phenomenal accents [404], [497] or salience [148], [529]. These usually
correspond to note starts, though not uniquelyit is possible that note ends
or changes in intensity can indicate beat, too. It is generally assumed that
stresses fall on the beat more often than not and that significant chordal
changes also do so. While this is not always the case, and indeed many musical
styles exhibit syncopation^ where there are off"-beat stresses, Steedman notes,
'No event inconsistent with either key or metre will occur in a piece until
sufficient framework (of key or time signature) has been established for it to
be obvious that it is inconsistent' [607]. There are counter-examples to this
statement, but it holds in the main.
There is a large body of literature in the music psychology and neuro-
science fields on how humans perceive rhythm. In particular, there is some
literature on human tapping processes and the behaviour of musicians versus
non-musicians (e.g. [155]). However, as the aim of most audio beat trackers
is to return to the underlying score or performance intentions rather than
replicate the perceptions of a listener, the general psychology literature will
not be discussed in detail here.

4.4 Onset Detection

While the metre and tempo of a piece of music can be thought of as a con-
stantly evolving signals, the musical events which underpin this are the starts
of notes, and these are discrete events. Many methods for beat tracking deal
with symbolic or MIDI data which represent these note start (onset) times.
It is highly possible, and indeed common, to simply attach an onset detector
to find the note starts in an audio signal and then track the resulting set of
discrete impulses. When this approach is used, the success of any beat tracker
is dependent upon the reliability of the data which is provided as an input.
Thus, detecting note starts in the audio can be as important as the actual
beat-tracking algorithm.
Note ends, even when played exactly as written in the score, can be ig-
nored as unreliable indications of beat due to reverberation, sustain or at the
opposite extreme, staccato events, where the note is cut short.
Note sources generally fall into two categories: harmonic and percussive.
The former produce sounds which would be regarded as notes, have an identi-
fiable pitch and harmonically related partials. Percussive sounds, in compari-
son, are more analogous to noise clouds. Drums and cymbals are the obvious
examples of this class. It should be noted that many (indeed most) pitched
108 Stephen Hainsworth

instruments have a transient onset which has much in common with percussive
sounds. Percussive sounds are usually characterized by significant increases in
signal energy (a 'transient') and methods for detecting this type of musical
sound are relatively well developed. Harmonic change with httle associated
energy variation is much harder to reliably detect and has received less atten-
tion in the literature. Two recent studies of onset detection are Bello et al. [30]
and Collins [95].
While the discussion below assumes that a hard detection decision is made
as to whether an onset is present at a given location, the beat trackers dis-
cussed below which work on continuous detection functions also need to trans-
form the raw audio into something more amenable. They also process the
signal in ways similar to those described below but do not perform the step
of making hard onset detection decisions, instead leaving this to the later
beat-tracking process. The hard-decision onset detection method yields a set
of discrete onset times, whereas the latter method results in a continuous
function from which beat tracking is performed.

4.4.1 Transient Event Detection

Transient events, such as drum sounds or the start of notes with a signifi-
cant energy change (e.g. piano, guitar), are easily detected by examining the
signal envelope. A typical approach, which is an adaptation of methods used
by a variety of other researchers [148], [392], [564], proceeds as follows: An
energy envelope function E^ (t) is formed by summing the power of frequency
components in the spectrogram for each time slice over the range required:

E,(n)=5]|STFT-(n,A:)p, (4.1)
kGKj

where STFT^(A:,n) is the short-time Fourier transform (STFT) of the signal

x{n) with rectangular window w centred at time n; fc is the frequency index
(see Chapter 2 for details). Usually analysis frames of about 20 ms are used
in computing the energy envelope, with 50-75% overlap between successive
frames. Different bands j can be used; for instance, low frequency information
covering the range 20-200 Hz is useful to separate. Setting KJ to the middle
range of 200 Hz to 15 kHz covers the majority of the harmonic information;
meanwhile, extending over 15-22.05 kHz (assuming a sample rate of 44.1 kHz),
the upper band is often generally free from harmonic content but contains a
clear indication of any strong transient information [444]. This is contrary to
the opinion of Duxbury [164], who claimed that there is no useful information
in this range. Many other ways to split the frequency spectrum have also been
proposed. One common approach is to use 5-10 sub-bands that are distributed
uniformly on a logarithmic frequency scale.
Ej{n) is not an ideal signal representation for detecting onsets. A potential
approach for improving it uses a three-point linear regression to find Dj(n),
 4 Beat Tracking and Musical Metre Analysis 109

the gradient of Ej(n), and peaks in this function are detected. The linear
regression fits a fine Yi = a -\- hXi + e^ to a set of N data pairs; we are
only interested in the estimate of h which is given by 6 = (^2=1 ^i^i
^ ^ ^ ) / ( E i I i ^i - NX'^), where X and Y denote the average of X and F ,
respectively. In the case here, X is the equi-spaced set of time indices n in
Ej(n) and Y is the corresponding E j . In the case where N = 3, this reduces
to
D,(n) = E > + 1 ) - E , ( n - 1 ) _ ^^_2^
O

It should be noted that the commonly used technique of differencing the signal,
where T)j{n) = Ej{n) Ej(n 1), is simply linear regression with A^ set to 2.
The linear regression approach, like that of Klapuri [347], aims to detect the
start of the transient, rather than the moment it reaches its peak power.
Dj(n) is often called a detection function [30] and is a transformed and
reduced signal representation. Subsequent processing needs to detect the on-
sets contained within this. This is usually done by simply selecting maxima
in Dj(n) and discarding peaks which do not pass a series of tests. Low-energy
peaks should be ignored (for instance by testing if they are less than two times
the local 1.5-second average of Ej) and peaks can also be ignored if there is
a higher-energy peak in the local vicinity'^ by using Dixon's timing criterion
[148]. Thresholds and constants are usually heuristically determined and de-
signed to give reasonable performance with a large range of styles. Figure 4.2
shows an example of a peak extraction method. When several sub-bands j
are involved, the functions Dj(n) can be combined by half-wave rectifying
and across-band summing before the peak-picking process [37], [347].

4.4.2 Pitched Event Detection

Detection of note starts where there is no associated energy transient (e.g.

violins, choral music) has received less attention than the easier problem ad-
dressed above. Notable recent exceptions are Klapuri [349], who used very
narrow frequency bands to detect changes in frequency; Laurent et al. [395],
who used wavelets; Davy and Godsill [123], who took a support vector ma-
chine approach; Desobry et al. [143], who furthered Davy's research and also
used kernel methods; and Abdallah and Plumbley [1], who used independent
component analysis (ICA) to generate a 'surprise' measure followed by an
HMM to perform reliable detection. Also, Bello et al. [31] utilized phase in-
consistencies in a manner very similar to time reassignment and Duxbury
et al. [164] proposed a spectral change distance measure adapted from the
Euclidean measure which was then applied to adjacent spectrogram frames.
Recently, Duxbury, Bello et al. [162], [163] have combined the previous two
approaches into a single measure for detection of harmonic changes via either

^This is similar to the psychoacoustic masking thresholds found for humans [475],
[694].
110 Stephen Hainsworth

1500

0.6 h
S ^ 0.4h
S 0-2 h
-B Of'ti
O -0.2 h
-0.4 ^
l+lhlWfr HWH^ x J rM/'|if*k^

500 1000 1500

Time (frames)

Fig. 4.2. Example of onset detection for transient events. The upper plot shows the
energy-based detection function, Ej(n); also shown are horizontal lines giving the
1.5 s local average of the energy function and x's showing the detected onsets. The
lower plot shows the gradient function Dj (n) from which peaks are found.

or both of phase inconsistency or spectral change. This method shows im-

provements over both individual approaches.
An alternative proposed by Hainsworth and Macleod [265] is the so-called
modified KuUback-Leibler distance measure given by

^ /M , f |STFT:(n,fc)| ^
(4.3)

(4.4)
fcG/C,d(fc)>0

where STFT^(n, k) is the STFT computed with window w. The measure em-
phasizes positive energy change between successive frames and /C defines the
spectral range over which the distance is evaluated (30 Hz to 5 kHz is sug-
gested as it represents the majority of clear harmonic information in the spec-
trum). Another advantage of this method is that it also takes into account
any transient energy which happens to be present as a useful aid.
A window length of about 90 ms is sufficient to give good spectral reso-
lution. To overcome frame to frame variation, histogramming of five frames
(weighted backwards and forwards with a triangular function) before and after
the potential change point was used and also a very short frame hop length
(namely, 87.5% overlap) was chosen to increase time resolution.
 4 Beat Tracking and Musical Metre Analysis 111

Detection of the peaks in this measure is a separate problem and is dis-

cussed more fully in [265]. Figure 4.3 gives an example of detection using this
method.

50001

45001

40001

35001
-''^*SWH*-^^ifi*** ,^
^30001 ^^*^^:ifi:iiiiii|

i 2500|
Plpll
2 2000I

15001
- '.*^ ^i^^
iooof IwiiiMWi hppi
mik;^ . t ^

>> " ' ' m , ? '4^ *'^TT*^*^^'^

mwi. w 1 " J J J ' " ^ 1 ^_^^
500 E
r'-*H*i
MSMSIS ^5
6 10
Time (s)

Fig. 4.3. Example of the output from the MKL harmonic change detection measure
for an excerpt of Byrd's 4-Part Mass. Onsets were missed at 1 s and 5.9 s while the
onset at 8.1 s is mis-estimated and should occur about 0.1 s later.

4.5 Rule-Based Approaches

We shall now discuss a number of broad methodologies for beat tracking in
turn. Rule-based approaches were among the earliest used when computers
were not capable of running complex algorithms. They tend to be simple and
encode sensible music-theoretic rules. Tests were often done by hand and were
limited to short examples. Often these did not even have expressive timing
added to them and only aimed to extract the most likely pulse given the
rhythmic pattern and tempo.
Steedman [607] produced one of the earliest computational models for
rhythmic analysis of music. His input was symbolic and with a combination
of musical structure recognition (especially melodic repetition) and psycho-
logically motivated processing, he attempted to parse the rhythmic structure
of Bach's 'Well Tempered Clavier' set of melodies. Similarly, Longuet-Higgins
and Lee [418] proposed a series of psychologically motivated rules for finding
112 Stephen Hainsworth

the beat and higher metrical levels from lists of onset times in a monophonic
melody. The rules were never implemented by the authors in the original pa-
per for more than five-bar examples, though there have since been several
papers by Lee which are summarized by Desain and Honing [141].
Parncutt [497] developed a detailed model for salience or phenomenal ac-
cent, as he termed it, and used this to inform a beat induction algorithm. Also,
he modelled medium tempo preference explicitly and combined these two in
a model to predict the tactus for a series of repeated rhythms played at dif-
ferent speeds. Comparison to human preferences was good. Parncutt's focus
was similar to that of Povel and Essens [529], while Eck [165] also produced
a rule-based model which he compared to Povel and Essens and others.
Temperley and Sleator [622] also used a series of rules to parse MIDI
streams for beat structure. They quoted Lerdahl and Jackendoff's generative
theory of tonal music (GTTM) [404] as the starting point of their analysis,
using the GTTM event rule (align beats with event onsets) and length rule
(longer notes ahgned with strong beats). Other rules such as regularity and a
number based on harmonic content were also bought into play. The aim was
to produce a full beat structure from the expressive MIDI input, and a good
amount of success was achieved.^

4.6 Autocorrelation Methods

Autocorrelation is a method for finding periodicities in data and has hence
been used in several studies. Without subsequent processing, it can only find
tempo and not the beat phase.
The basic approach is to define an energy function E(n) to which local
autocorrelation is then applied (in frames of length Tw, centred at time n):
Tw/2
r(n,i)= ^ E{n-\-u)E{n + u-i). (4.5)
n=-(Tw/2) + l

The value of i which maximizes r(n, i) should correspond to the period-length

of a metrical level. This will often be the beat, but it is possible that if the
tatum is strong that autocorrelation will pick this instead.
Tzanetakis et al. [631], [633] included a series of rhythmic features in their
algorithm for classification of musical genre. While not specifically extracting
a beat, it performs a function similar to beat analysis. Their method was
based upon the wavelet transform, followed by rectification, normalization,
and summation over different bands before using autocorrelation to extract
periodicity. Local autocorrelation functions were then histogrammed over the
entire piece to extract a set of features for further use; these tend to show
more coherence for rock pieces than for classical music.
^Source code for Temperley's method is available in [596].
 4 Beat Tracking and Musical Metre Analysis 113

Foote and Uchihashi [194] used the principle of audio self-similarity to

examine rhythmic structure. The assumption was that within the space of a
single sub-beat, the sound is approximately constant and therefore the spec-
trum will have high similarity. They therefore defined a similarity measure
as the normalized scalar product (computed over the frequencies k) of the
magnitude spectra of frames at times Ui and rij

where w is some window. This produced a two-dimensional plot of similarity

between any two frames of the audio signal, which was then autocorrelatively
analysed for tempo hypotheses using

B{ni,nj) = ^dpooteirii' ,nf)dpooteini + Hi' ,nj + n j / ) . (4.7)

This was extended to be time varying, hence producing their 'beat spectro-
gram', which was a plot of the local tempo hypothesis versus time.
Other autocorrelation approaches include Mayor [445], who presented a
somewhat heuristic approach to audio beat tracking: a simple multiple hy-
pothesis algorithm was maintained which operated on his so-called BPM
spectrogram, BPM referring to beats per minute. Also Paulus and Klapuri's
method [503] for audio beat analysis utilized an autocorrelation-like function
(based on de Cheveigne's fundamental frequency estimation algorithm [135]),
which was then Fourier transformed to find the tatum. Higher-level metri-
cal structures were inferred with probability distributions based on accent
information derived using the tatum level. This was then used as part of an
algorithm to measure the similarity of acoustic rhythmic patterns. Brown [55]
used her narrowed autocorrelation method to examine the pulse in musical
scores. Davies and Plumbley [118] and Alonso et al. [15], [16] have also pro-
duced autocorrelation-based beat trackers.

4.7 Oscillating Filter Approaches

There are two distinct approaches using oscillating filters: In the first, an
adaptive oscillator is excited by an input signal and, hopefully, the oscillator
will resonate at the frequency of the beat. The second method uses a bank of
resonators at fixed frequencies which are exposed to the signal and the filter
with the maximum response is picked for the tempo. Beat location can be cal-
culated by examining the phase of the oscillator. This method is particularly
suited to causal analysis.
The first, single-filter approach is typified by Large [389], [390], who used a
single non-linear oscillator with adaptive parameters for the phase, frequency,
and update rate, though these were initialized to the correct settings by hand.
114 Stephen Hainsworth

The observed signal is a set of impulses s{n) = 1 when there is an onset event
and s{n) = 0 otherwise. The oscillator is given by

o(n) = 1 -h tanha(cos27r(/>(n) 1), (4.8)

where o{n) defines an output waveform with pulses at beat locations with
width tuned by a; see Fig. 4.4. The phase is given by

4>{n) = n rii (4.9)

P
where rii is the location of the previous beat and p is the period of oscillation
(tempo). Crucially, the single oscillator in (4.8) is assumed not to have a
fixed period or phase and updates are calculated every time an onset event is
observed in s{n) using

Arii = r]is{n)^sech^{a{cos27r(j){n) - I)}sin27r0(n), (4.10)

Ap = rj2s{n)^sech^{a{cos27r(l){n) I)}sin27r0(n), (4.11)

where rji and r]2 are 'coupling strength' parameters. The update equations
enable the estimation of the unknown parameters p and rii. Marolt [433],
however, points out that oscillators can be relatively slow to converge because
they adapt only once per observation.

a = 1 , p = 10 a=10,p = 10 a= 1, p = 5

0'^ i
^^\ /^'^ ^'^\ /-^'^ ^^ QL
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
timet time t time t

Fig. 4.4. Example output signals o{t) generated using (4.8) for various values of a
and p.

Large's test data was a series of impulses derived from expressive MIDI
performances and the aim was to track the pulse through the example. An
extra level of complexity which allowed the system to continue following the
beat was to have a second oscillator 180 out of phase which could take over
control from the first if confidence dropped below a certain threshold.
McAuley [450] presented a similar adaptive oscillator model to that of
Large and indeed compared and contrasted the two models. Similarly, Toivi-
ainen [626] extended Large's model to have short- and long-term adaption
mechanisms. The former was designed to cope with local timing deviations
while the latter followed tempo changes. It was tested on expressive MIDI
 4 Beat Tracking and Musical Metre Analysis 115

performances. Another variation is that of Eck [166], who used Fitzhugh-

Nagumo oscillators (models of neural action) linked by Heaviside coupling
functions into networks. His focus was to reproduce the downbeat extraction
of Povel and Essens [529] from synthetic onset data. Various authors [166],
[389] have also suggested that adaptive filters have neurological plausibihty
and this is their motivation for its use.
The second approach is typified by Scheirer [564], who produced one of the
first systems for beat tracking of musical audio. The difference compared with
Large's method is that Scheirer's method implemented a bank of comb filters
at different fixed feedback delays and searched for the one which resonated
best with the input signal at any given time. It should be noted that the
bank responds in a comb-like manner with multiples and subdivisions of the
tempo also showing resonation to the signal. Scheirer implemented 150 filters
logarithmically spaced between 60 bpm and 240 bpm, where bpm stands for
'beats per minute'. The input audio signal was treated in six sub-bands to find
rectified power envelopes as a function of time. Each sub-band was processed
by a separate comb-filter bank before the outputs were summed and the oscil-
lator with the greatest response picked as the current tempo. Phase was also
considered so as to generate a tapping signal corresponding to the tactus.
The model worked with considerable success, although there remained the
problem of a 2-3 second burn-in period needed to stabilize the filters, and also
a propensity for the algorithm to switch between tracking the tactus and its
subdivisions/multiples since Scheirer did not explicitly address the stability
of the beat estimate. Klapuri [349] (see below) capitalized on the latter ob-
servation in his method, using a bank of comb-filter resonators as the initial
processing method for his system. McKinney and Moelants [452] also found
a resonator method for tempo extraction to outperform histogramming and
autocorrelation approaches.

4.8 Histogramming Methods

Several approaches have focused on audio beat tracking using histogram-
ming of inter-onset intervals. First, the signal is analysed to extract onsets
before the subsequent processing takes place. This was discussed above in
Section 4.4. Differences between successive onsets can be used (first-order
intervals), though it is more productive to also use the differences between
onsets that are further apart (all-order intervals). The motivation for this is
that often the successive onsets define the tatum pulse rather than the tactus,
which can be better found using onsets spaced further apart. Histogramming
has similarities to the autocorrelation approaches of Section 4.6, though with
a discrete input rather than the continuous signal used for autocorrelation.
There are various methods of performing the histogramming operation;
defining the set of calculated inter-onset intervals (lOIs), denoted Oi, i =
1,2,..., one can follow Seppanen [572] and divide the lOI time axis into J
116 Stephen Hainsworth

bins and count the number of lOIs which fall in each: h{j) = count(i, \oi
u{j-\-0.5)\ < O.bu) where u is the width of a bin. In contrast, Gouyon
et al. [247] and Hainsworth [263] treat the lOI data as a set of Dirac delta
functions and convolve this with a suitable shape function (e.g. a Gaussian).
The resulting function generates a smoothly varying histogram. This is defined
as h{j) = Xli^i *-^(i)5 where * denotes convolution and Af{j) is a suitable
Gaussian function (low variance is desirable). Peaks can then be identified and
the maximum taken as the tempo. Alternatively, Dixon [148] gives pseudocode
for an lOI histogram clustering scheme.
Seppanen [572] produced an archetypal histogramming method. After an
onset detection stage, he first extracted tatums via an inter-onset interval his-
togramming method. He then extracted a large number of features (intended
to measure the musical onset salience) with the tatum signal informing the
locations for analysis. These features were then used as the input to an al-
gorithm based on pattern recognition techniques to derive higher metrical
levels including the pulse and bar lines. Seppanen [573] gives further details
of the tatum analysis part of the algorithm. The final thing to note is that
the method was the first to be tested on a statistically significant audio data-
base (around three hundred examples, with an average length of about one
minute).
Gouyon et al. [247] applied a process of onset detection to musical audio
followed by inter-onset interval histogramming to produce a beat spectrum.
The highest peak (which invariably corresponded to the tatum) was then
chosen as the 'tick'. This was then used to attempt drum sound labelling
in audio signals consisting solely of drums [245], to modify the amount of
swing in audio samples [244], and to investigate reliable measures for higher
beat level discrimination (i.e., to determine whether the beat divided into
groups of two or three) [246]. Other histogramming methods include Wang
and Vilermo [661], Uhle and Herre [635], and Jensen and Andersen [318], all
of which present variations on the general approach and use the results for
different applications.

4.9 Multiple Agent Approaches

Multiple agent methods are a computer science architecture. While there is

a great deal of variation in the actual implementation and often the finer
details are left unreported, the basic philosophy is to have a number of agents
or hypotheses which track independently; these maintain an expectation of
the underlying beat process and are scored with their match to the data.
Low-scoring agents are killed while high-scoring ones may be branched to
cover differing local hypotheses. At the end of the signal, the agent with the
highest score wins and is chosen. Older multiple agent architectures include
the influential model of Allen and Dannenberg [14] and Rosenthal [547]. The
 4 Beat Tracking and Musical Metre Analysis 117

two most notable multiple agent architectures are those of Goto and that of
Dixon.
Goto has produced a number of papers on audio beat tracking of which
[221], [238], [240] are a good summary. His first method centred on a multi-
ple agent architecture where there were fourteen transient onset finders with
slightly varying parameters, each of which fed a pair of tempo hypothesis
agents (one of which was at double the tempo of the other). A manager
then selected the most reliable pulse hypothesis as the tempo at that in-
stant, thereby making the algorithm causal. Expected drum patterns as a
strong prior source of information were used and tempo was tracked at one
sub-beat level (twice the speed) as well as the pulse in order to increase
robustness.
This method worked well for audio signals with drums but failed on other
types of music. Thus, he expanded the original scheme to include chord change
detection [240], each hypothesis maintaining a separate segmentation scheme
and comparing chords before and after a beat boundary.
Dixon [148] has also investigated beat tracking both for MIDI and audio,
with the aim of out putting a sequence of beat times. The algorithm performed
well with a MIDI input, and with the addition of an energy envelope onset
detection algorithm, it could also be used for audio (though with lower perfor-
mance). The approach was based upon maintaining a number of hypotheses
which extended themselves by predicting beat times using the past tempo tra-
jectory, scored themselves on musical salience, and updated the (local) tempo
estimate given the latest observation. The tempo update was a function of
the time coherence of the onset, while the salience measure included pitch
and chord functions where the MIDI data was available. Hypotheses could be
branched if onsets fell inside an outer window of tolerance, the new hypothesis
assuming that the onset was erroneous and maintaining an unadjusted tempo.
Initialization was by analysis of the inter-onset interval histogram. Dixon has
also used his beat tracker to aid the classification of ballroom dance samples
by extracting rhythmic profiles [149].

4.10 Probabilistic Models

Probabilistic approaches can have similarities to multiple agent architectures

in that the models underlying each can be very similar. However, while the
latter use a number of discrete agents which assess themselves in isolation,
probabihstic models maintain distributions of all parameters and use these
to arrive at the best hypothesis. Thus, there is an explicit, underlying model
specified for the rhythm process, the parameters of which are then estimated
by the algorithm. This allows the use of standard estimation procedures such
as the Kalman filter [41], Markov chain Monte Carlo (MCMC) methods [208],
or sequential Monte Carlo (particle filtering) algorithms [22] (see Chapter 2
for an overview of these methods).
118 Stephen Hainsworth

This section will concentrate on some of the models developed rather than
details of the estimation procedures which are used to evaluate the final an-
swer, as these can often be interchangeable (a point made by Cemgil, who
used a variety of estimation algorithms with the same model [77]).
Again, the various methods can be broken down into two general groups:
those that work with a set of MIDI onsets (or equivalently a set of onsets
extracted from an audio sample) and those that work to directly model a
continuous detection function^ computed from the original signal.

4.10.1 Discrete Onset Models

Those who have worked on the problem include Cemgil et al. [77], who worked
with MIDI signals, and Hainsworth [263], who used Cemgil's algorithm as a
starting point for use with audio signals.
The crux of the method is to define a model for the sequential update of
a tempo process. This is evaluated at discrete intervals which correspond to
note onsets. The tempo process has two elements: the first defines the tempo
and phase of the beat process. The second is a random process which proposes
notations for the rhythm given the tempo and phase. A simple example of this
is that, given a tempo, the time between onsets could either be notated as a
quaver or a crotchet, one speeding the tempo up and the other requiring it to
slow down. The probabilistic model will propose both and see which is more
hkely, given the past data (and future if allowed).
The model naturally falls into the framework for jump-Markov linear sys-
tems where the basic equations for update of the beat process are given
by

(7n)^n-l+V, (4.12)
S = UnOn + e. (4.13)

{s} is the set of observed onset times, while On is the tempo process at
iteration (observed onset) n and can be expanded as

Pn
On (4.14)
An

Pn is the predicted time of the n^^ observation Sn, and A^ is the beat period in
seconds, i.e. A^ 60/pri where p^ is the tempo in beats per minute. ^^(Tn)
is the state update matrix, H^ = [ 1 0 ] is the observation model matrix, and
Vn and Cn are noise terms; these will be described in turn.
The principal problem is one of quantizationdeciding to which beat or
sub-beat in the score an onset should be assigned. To solve this, the idealized

Strictly speaking, it will be pseudo-continuous due to sampling.

 4 Beat Tracking and Musical Metre Analysis 119

r r r ^
^ p
Fig. 4.5. Figure showing two identical isochronous rhythms. The top rhythm is
much more likely in a musical notation context than the lower.

(quantized) number of beats between onsets is encoded as the random jump

parameter, 7^, in <Pn(7n),

1 In (4.15)
^n(7n)
0 1
In Cn-l' (4.16)
While the state transition matrix is dependent upon 7^, this is a differ-
ence term between two absolute locations, Cn and Cn-i- Cn is the unknown
quantized number of beats between the start of the sample and the n*^ ob-
served onset. It is this absolute location which is important and the prior on
Cn becomes critical in determining the performance characteristics. This can
be elucidated by considering a simple isochronous set of onsetsif absolute
score location is unimportant, then the model has no way of preferring aligning
them to be on the beat over placing them on, say, the first semiquaver of each
beat. This is demonstrated in Fig. 4.5. Cemgil [77] broke a single beat into
subdivisions of two and used a prior related to the number of significant digits
in the binary expansion of the quantized location. In MIDI signals there are no
spurious onset observations and the onset times are accurate. In audio signals,
however, the event detection process introduces errors both in localization ac-
curacy and in generating completely spurious events. Thus, Cemgil's prior is
not rich enough; also, it cannot cope with compound time, triplet figures, or
swing. To overcome this, Hainsworth [263] broke down notated beats into 24
sub-beat locations, c^ = {1/24, 2 / 2 4 , . . . , 24/24, 25/24,...}, and a prior was
assigned to the fractional part of c^,

p{cn) ocexp(-Alog2{c^}), (4.17)

where c^ is the denominator of the fraction of Cn when expressed in its most
reduced form; i.e., (i(3/24) = 8, d(36/24) = 2, etc. A is a scale parameter
determining the sensitivity of the prior. This is shown graphically in Fig. 4.6.
The prior is improper (i.e., it does not sum to unity), which is why p{cn) is
only expressed as a proportionality. The integer part of Cn increases as the
number of beats processed increases. As a result of this, jn is always strictly
positive; it will be less than 1 if a sub-beat interval is observed, but if there
is more than one beat between observed, onsets, 7n will be greater than 1.
120 Stephen Hainsworth

018
Jl
016

0.14 -
012

0.1
3

J)
^ 1 -
-
0.08
} -
006 J -
004 -
002 -
n
15
c,,

Fig. 4.6. Graphical description of the prior upon Cn- The horizontal axis is the
sub-beat location from 1 to 24, while the associated probability p(cn) is shown on
the vertical axis.

quantized score locations

note lengths

tempo process

observations

Fig. 4.7. Directed acyclic graph of the jump-Markov linear system beat model.
The dependence between Cn and jn is deterministic, while other dependencies are
stochastic.

The tempo process has an initial prior, p{6o), associated with it. For the
purposes of a general beat-tracking algorithm, it is assumed that the likely
tempo range is 60 bpm to 200 bpm and that the prior is uniform within this
range.
So far, the model for tempo evolution and proposing a set of onset times
has been considered. Finally, the observation model must be specified. Sn is
the n^^ observed onset time and therefore corresponds to the pn in On- Thus,
Hn = [l O]- The state evolution error, v^, and observation error, e^, are
given suitable distributionsusually for mathematical convenience, these are
zero-mean Gaussians with appropriate covariances [26]. The overall model can
be summarized by a directed acyclic graph (DAG) as shown in Fig. 4.7. It
should be noted that even spurious onsets are assigned a score location.
 4 Beat Tracking and Musical Metre Analysis 121

When working with real-world audio signals, more information than just
the onset times can be extracted from the signal, and this can aid the analysis
of the rhythm. The most obvious example is the amplitude of onsets while
others include a measure of chordal change and other 'salience' features as
postulated by Parncutt [497] and Lerdahl and Jackendoff [404]. Hainsworth
[263] utilized these in his model as a separate jump-Markov linear system for
amplitude and a zero-order Markov model for salience (here, the salience is
only a function of the current state and has no sequential dependency). There
has been little research into appropriate measures of salience for extracting
accents in music; other than the papers mentioned above, Seppanen [573] and
Klapuri [349] also proposed features which perform this function.
Given the above system, various estimation procedures exist. Cemgil [77]
described the implementation of MCMC methods as well as particle filters to
estimate the maximum a posteriori (MAP) estimate for the rhythm process,
while Hainsworth [263] utilized particle filters to find the MAP estimate for the
posterior of interest given hy p{ci:m Oi:n, (^i:n\si:m CLi-.m Si.n)^ where ai:n was
the underlying amplitude process observed as ai:n, and Si.n was the observed
set of saliences. Full details can be found in either of the publications.
Other similar methods include an earlier approach of Cemgil's [79] where
what he termed the 'tempogram' (which convolved a Gaussian function with
the onset time vector and then used a localized tempo basis-function*^ to
extract a measure of tempo strength over time) was tracked with a Kalman
filter [41] to find the path of maximum tempo smoothness.
Raphael's methods [537] were based around hidden Markov models where
a triple-layered dependency structure was used: quantized beat locations in-
formed a tempo process which in turn informed an observation layer. The
Markov transitions were learned between states from training data, and then
the rhythmic parse evaluated in a sequential manner to decide which was the
most likely tempo/beat hypothesis. This was tested on both MIDI and au-
dio (after onset detection) and success was good on the limited number of
examples, though manual correction from time to time was permitted.
Laroche [392], [393] used a maximum likelihood framework to search for
the set of tempo parameters which best fit an audio data sample. The in-
put was processed by typical energy envelope difference methods to extract
a list of onset times. Inter-onset times (which are phase independent) were
then used to provide likelihoods for the 2-D search space with discretized
tempo and swing as the two axes. This algorithm has been included in com-
mercially available Creative sound modules for several years. Lang and de
Preitas [387] presented a very similar algorithm to that of Laroche but used
a continuous signal representation and a slightly more complex estimation
procedure.

^The tempo basis function was defined as a set of weighted Dirac functions
V^(t;r, a;) = SJ^_oo ^*^^+*2'^(^) ^^ ^ delay of r and spaced with frequency (and
hence tempo) given by uj.
122 Stephen Hainsworth

Hainsworth also presented a second algorithm which is essentially a refor-

mulation of the above but using Brownian motion relations as a base [266]. It
was not as successful as the above model. Others include Takeda et al. [617]
and Lam and Godsill [386].

4.10.2 Continuous Signal Representations

The second approach to tracking the beat with stochastic models uses a detec-
tion function and attempts to model this directly instead of extracting onsets
first. As such it must have all the elements of the above models, including a
tempo process and a model for the likelihood of an onset being present at any
given beat or sub-beat location; however, it must also have a model for the
signal itself and what is expected at an onset and between these.
Hainsworth [263] proposed a method using particle filters whereby the
tempo was modelled as a constant velocity process similar to the one described
above and which proposed onsets in a generative manner at likely sub-beat
locations. The signal detection function modelled was a differenced energy
waveform, utilizing high-frequency information, very similar to T)j{t) shown
in the lower plot of Fig. 4.2.
Onset locations can clearly be seen in this signal representation, and on
close examination all onsets have a very similar evolution in time which can
be well modelled by a hidden Markov model (HMM; see Chapter 2 for a
definition). This is performing the task of onset detection. The model used is
shown in Fig. 4.8 with each state having a different output distribution (also
termed likelihood). For mathematical convenience, these are Gaussians with
differing means and variances but sufficiently separated so that the output
distribution of state ^i does not significantly overlap with that of ^o or 5*2,
etc. This defines a generative model for the signalby generative, it is meant
that by using a random number generator and the specified distributions,
a process with the same statistical properties as the original signal can be
generated.
A naive scheme simply generates proposals from the prior distributions,
but the Viterbi algorithm (see [654] and Chapter 2) can be used to find the
best path through the HMM and also its probability, which simplifies the
calculation needed once an onset is hypothesized. The model worked well
on the small number of examples tried but required the expected sub-beat
structure to be specified by hand for robust performance.
In comparison, Sethares et al. [577] proposed four filtered signals (time do-
main energy, spectral centroid, spectral dispersion, and one looking at group
delay) which were then simply modelled as Gaussian noise with a higher vari-
ance at beat locations compared to between them. Looking back at Fig. 4.2,
it can clearly be seen where the variance of the generative noise process used
to model the signal would be higher. A model similar to those above was used
and a particle filter environment chosen for the estimation procedure. The
 4 Beat Tracking and Musical Metre Analysis 123

(a) Data with states superimposed.

(b) Directed acyclic graph of HMM model.

Fig. 4.8. HMM for beat-tracking algorithm with Viterbi decoding included. States
55, 56, and 5? are functionally equivalent to 54, and 58 is equivalent to So- The
null state, 59, has no observation associated with it, therefore making transition to
it highly unattractive.

model did not explicitly include a model for sub-beats but seemed to function
well on the data presented.
A somewhat different method for tracking the beat through music was
presented by Klapuri et al. [348], [349]. A four-dimensional observation vector
(as a function of time) was generated by applying a similar method to that
of Scheirer [564] to generate resonator outputs but using different frequency
bands and a different method for extracting the energy signal which also cap-
tures harmonic onsets. A measure of salience, dependent upon the normalized
instantaneous energies of the comb-filter resonators, was also attached to this.
A problem with Scheirer's method was that it was prone to switch be-
tween different tempo hypotheses (usually doubling or halving), and Klapuri
addressed this using an HMM to impose some smoothness to the tempo evo-
lution. He proposed a joint density for the estimation of the period-lengths
of the tatum, tactus, and measure level processes, applying a combination of
sensible priors and dependencies learned from data. The phase of the tatum
124 Stephen Hainsworth

and tact us pulse were estimated to maximize the observed salience at beats. In
estimating the phase of the super-beat (measure) structure, a key assumption
made was the expectation of two simple beat patterns which occur frequently
in so-called 4/4 time. While this should considerably aid performance with
music in this time signature, performance in the super-beat estimation was
degraded for examples with a ternary metre (e.g. 3/4). Nevertheless, the algo-
rithm was tested on a significant database and was successful. A comparison
is presented below.

4.11 Comparison of Algorithms

If the focus is restricted to beat tracking in musical audio signals, then the
methods discussed above in Sections 4.5 to 4.10 have various strengths and
weaknesses. This section will highlight them and then present a comparison
of several methods.
Rule-based approaches have never been applied to audio and have solely
been used to code sensible but simple music theoretic rules in order to model
music psychology expectations. The reason that they have never been used on
audio signals is possibly because they are not easily expanded to cope with
erroneous data and hence would perform poorly on the inexact data produced
by onset detection algorithms.
Autocorrelative and histogram methods have much in common; they are
both methods of obtaining a tempo profile, the difference being that autocor-
relation works with a sampled signal while histogramming works with discrete
onset times. They are therefore useful for finding the tempo but are not im-
mediately applicable to extracting the beat phase (this is a secondary task).
Adaptive oscillators are particularly suited to causal operation and have
some psychoacoustic justification [390]. However, they have not been applied
to audio signals. This may be because the update routines required on adaptive
single filters are not easily adaptable to real data or possibly because they are
not well able to cope with sub-beats. Many of the systems also required manual
initiation to set the correct tempo and phase. Comb-filters as implemented
by Scheirer [564] and used by Klapuri et al. [349] have been applied to audio
signals.
This leaves multiple agent approaches and probabilistic, model-based
methods. These two bear some significant similarities, but the latter delim-
its the underlying assumptions from estimation procedures whereas they are
intermixed by multiple agent methods. This makes the adaption and opti-
mization of the probabilistic models easier, though reasonable success has
been reported with both approaches.

4.11.1 Tests
There has been a move in recent years towards testing algorithms with a large
database of audio samples collated from all genres and usually from standard,
 4 Beat Tracking and Musical Metre Analysis 125

commercially available sources. This was begun with Seppanen [572] with a
database of 330 audio samples, while Klapuri [349] used 478. A comparison
was also undertaken by Gouyon et al. [248] into tempo induction from audio
signals using a large dataset of 3199 examples from three databases and is
currently the most extensive.
The comparison below used a hand-labelled database of 222 samples of
around one minute divided into six categories: rock/pop, dance, jazz, folk,
classical, and choral. The tempos were limited to the range 60-200 bpm with
the exception of the choral samples. Several examples exhibited significant ru-
bato, 8 had a rallantando (slowing down), and 4 had a sudden tempo change.
Forty-two also had varying amounts of swing added. Full details of the data-
base can be found in [263].
Another problem is how to evaluate the performance of a beat-tracking
algorithm. As of this writing, no study has yet made a serious attempt to no-
tate the complete rhythm and idealized score locations of every onset present
in the audio sample;^ rather the assessment has been limited to 'tapping in
time' to the sample and producing an output of beat times that agrees with
those of trained human musicians.
Klapuri [349] gives two criteria, which are adopted here, to judge the per-
formance of an algorithm on a particular example. The first is 'continuous
length' (C-L), by which it is meant the longest continually correctly tracked
segment, expressed as a percentage of the whole. Thus, a single error in the
middle of a piece gives a C-L result of 50%. Another, looser criterion is sim-
ply the total percentage of the whole which is correctly tracked (defined as
'TOT' from now on). Here, both are expressed as percentages of the manually
detected beats which are correctly tracked, rather than of the time stretches
these represent. Using Klapuri's definitions once again, a beat is determined
to be correctly tracked if the phase is within 15% and the tempo period is
correct to within 10%.
Here, the trackers^ of Scheirer [564], Klapuri [349], and Hainsworth [263]
are compared and the results are shown in Table 4.2. The columns under 'Raw'
are base results according to the above criteria; however, it is sometimes found
that the beat tracker tracks something which is not the predefined beat but is
a plausible alternative. Usually, this is half the correct tempo (in the case of
fast samples) or double (for particularly slow examples). When swing is en-
countered, it is occasionally possible for the trackers to even track at one and a
half times the tempo (i.e., tracking three to every two correct beats). Doubling
or halving of tempo is psychologically plausible and hence acceptable; however
the errors encountered with swing are not. The second set of columns com-
pares results once doubling and halving of tempo are allowed. Performance
on individual genres is shown graphically in Fig. 4.9 for Hainsworth's and
Klapuri's algorithms.

^The closest is probably Goto and Muraoka [234].

^The beat trackers tested were all the original authors' own.
126 Stephen Hainsworth

Maximum length correct (% of total)

100
|.
80
1 ^, 1
~\\
\\
'
>
I'Folk
rciassical
Uazz
1 60
1 V.^ \\
o \\ i^horal
^ 40 -Dance
Rock/pop
1 *^
1 \
1\ '^' ^' \\ 1 \ s
\ \
\
\
-
20

50
I ^^^
100
i
150
': l'' >: K: 200
Total percentage correct
100
801-
' 6o^
. 1 1
1
1 ^\ \ \
I \
: 4ol 1 1
\ 1
\ 1
201 1 1

1 50
1 1

100 150 200

Song index

(a) Hainsworth's results

Maximum length correct (% of total)

lOOr
^^^ 11
80 ^ Hi \
Vv 1 1 :i< sical 1
"
' 60h \\ 1 ^

V Choral
40 H-Dance ' 1 0

Rock/pop
20 ^ Jazz 1 \ FollA'
1 1
50 100 150 200
Total percentage correct

80
60 M'^
- '
\ ^~"\
'^ \ '
\
\s
1'
1'
||
"
ii

40 - \ \
1; ! Y',". -
20
- 1
(3 50 100
\

Song index
1

\ \
50
,1 \'i^; 200

(b) Klapuri's results

F i g . 4.9. Graphical display of the results for Hainsworth's (top) and Klapuri's beat
tracker. The solid line is the raw result while the dashed line is the 'allowed' result.
Note that ordering is strictly by performance for each genre under any particular
criteria.
 4 Beat Tracking and Musical Metre Analysis 127

Table 4.2. Comparison of results on the database. The three beat trackers use
audio adata as inputs.
Raw Allowed
C-L (%) TOT (%) C-L (%) TOT (%)
Hainsworth 45.1 52.3 65.5 80.4
Scheirer 23.8 38.9 29.8 48.5
Klapuri 55.9 61.4 71.2 80.9

It can be seen that Klapuri's model performs the best in terms of raw
results and continuous tracking, while the performance of Hainsworth when
considering total number of beats with allowed tempo mistakes is about equiv-
alent. Klapuri's method performs better than Hainsworth's with rock/pop and
dance, though it fails somewhat with jazz. Hainsworth's outperforms Klapuri's
on choral music, probably because of the onset detection algorithm used by
Hainsworth (described above in Section 4.4), which gives superior performance
for these choral samples.
Both Klapuri's and Hainsworth's models significantly outperform Scheirer's.
Klapuri [349] compared his model to Scheirer's and also Dixon's [148] mod-
ified MIDI beat-tracker. Seppanen [572] reported that his program was less
successful than Scheirer's, tested on a large database that was a subset of
Klapuri's. Also, on the related issue of tempo induction, the comparison by
Gouyon et al. [243] showed that Klapuri's method performed the best at this
task.
Finally, performance of one of the stochastic models which uses a sig-
nal representation is shown on a single example in Fig. 4.10. This shows
Hainsworth's second stochastic model (described above in Section 4.10.2) with
a swing example. The model is very successful at extracting onsets and is
good at tempo tracking. The limitation is that the expected sub-beat struc-
ture has to be specified in advance. Thus, the model cannot be considered
pan-genre.

4.12 Conclusions

This chapter has discussed a number of differing approaches to the generic

task of 'beat tracking'. Under this catch-all term, there are actually a number
of possible goals, from replicating human tempo preference to a full labelling
of every onset as to its correct quantized score location. Recent methods have
aimed to extract the correct tempo and beat phase from audio signals ('tap-
ping in time to the music').
Current methods such as Klapuri's [349] or Hainsworth's [263], [266] are,
starting to achieve a reasonable level of success over databases of significant
size and complexity. However, they are less successful on certain genres such
as jazz (where part of the appeal of the style is its rhythmic complexity)
128 Stephen Hainsworth

10 20 Time (s) 30 40 50

(a) Tempo profile.

0.4

0.2

0
n\fwi^ |44w
-0.2

2 4Time(s) 6 8 10

(b) Onset detection process.

Fig. 4.10. Output of Hainsworth's second stochastic beat tracker (see Sec-
tion 4.10.2) for a swing example, a) shows tracked tempo (dashed) and hand-labelled
tempo (solid); b) shows the onset detection process for the first 10 seconds with solid
vertical lines denoting detected beats and dashed vertical lines showing the detected
swung quavers.

and classical music (which is prone to radical rhythmic evolution and also
has fewer easily extractable beat cues). Classical music particularly seems
to require pitch analysis in order to extract reliable beat cues. Thus, while
the aim is obviously to have a generic beat tracker which works equally well
with all genres, it is likely t h a t in the short term, style-specific cues will have
t o be added. Klapuri [353] and Goto [221] b o t h apply knowledge of typical
d r u m p a t t e r n s in popular music to their algorithms. Dixon [149] goes a step
further and uses rhythmic energy p a t t e r n s extracted from audio samples to
aid classification of ballroom dance examples, a process which could easily be
reversed to aid beat tracking.
In addition to better modelling specific styles and the rhythmic expecta-
tions therein, the second area for expansion is to look at better signal rep-
resentations for extracting the cues needed to perform beat tracking. Rock
and p o p music with its drum-heavy style is easily processed using energy
measures; classical music is much harder to process and only relatively re-
cently have methods been applied to extract note changes where there is little
transient energy. These will need to be improved.
 4 Beat Tracking and Musical Metre Analysis 129

In conclusion, the field of beat tracking or rhythmic analysis is one area

of musical audio processing where some significant success has been achieved
and there is much to build upon. However, there is also room for improvement
and further accomplishments.
Unpitched Percussion Transcription

Derry FitzGerald^ and Jouni Paulus^

^ Cork Institute of Technology, Rossa Avenue, Bishopstown, Cork, Ireland

derry.fitzgeraldOcit.ie
^ Institute of Signal Processing, Tampere University of Technology,
Korkeakoulunkatu 1, 33720 Tampere, Finland
j ouni.paulusQtut.f i

5.1 Introduction

Up until recently, work on automatic music transcription has concentrated

mainly on the transcription of pitched instruments, i.e., melodies. However,
during the past few years there has been a growing interest in the problem of
transcription of percussive instruments. This chapter aims to give an overview
of the methods used in this field ranging from the pioneering works of the 1980s
to more recent systems.
There currently exists a large family of percussive instruments, many of
which have been in existence for quite a long time [293]. The work presented in
this chapter concentrates on the transcription of unpitched^ percussive instru-
ments. In particular, the vast majority of research to date has focused on the
unpitched percussion instruments found in Western popular music, with a par-
ticular focus on the drums found in the standard rock/pop drum kit, namely
snare drum, kick drum (also known as a bass drum), tom-toms, hi-hats, and
cymbals. A notable exception to this is the work on tabla transcription by
Gillet and Richard [211]. However, many of these methods could be utiUzed
for transcribing other percussive sounds if suitable modifications are made.
The percussion instruments mentioned in this chapter can be divided into
two main types: membranophones and idiophones. Membranophones, includ-
ing drums such as snare drums, kick drums, and tom-toms, typically consist
of a membrane or skin stretched across a frame. Idiophones, including instru-
ments such as hi-hats and cymbals, are typically rigid bodies, such as a metal
plate. In both cases, sound is produced by striking the membrane or plate. In
a standard rock/pop drum kit, striking is usually done with a wooden drum-
stick, except for the kick drum which is struck using a beater made of epoxy or
rubber mounted on a foot pedal. The striking of a given drum can be modelled
"^The word 'unpitched' is used here to emphasize the fact that the instruments
are normally not used to play melodies, even though many drums can be tuned and
their sound evokes a perception of pitch.
132 Deny FitzGerald and Jouni Paulus

Kick drum Snare drum Crash cymbal

0.2 0.4 1 2 3
Time (s) Time (s)

Fig. 5.1. Example waveforms. The images on the top row are the time domain
waveforms of a kick drum, a snare drum, and a crash cymbal, from left to right. The
lower row contains the corresponding spectrograms in the same order. The sound
samples are from the RWC Musical Instrument Sound Database [230].

as an impulse function and so a broad range of frequencies will occur in the im-
pact. As a result, all possible modes of vibration of the plate or membrane are
excited simultaneously, and the narrower the frequency band associated with
a given mode, the longer it sounds. The interested reader is referred to [193]
for a mathematical discussion of the properties of ideal membranes and plates.
Many of the membranophones used in a standard rock/pop drum kit can
be tuned by adjusting the tension of the membrane. In conjunction with the
different sizes available for each drum type, this results in considerable varia-
tions in the timbre obtained within a given drum type. Nevertheless, it can be
noted that the membranophones have most of their spectral energy contained
in the lower regions of the frequency spectrum, typically below 1000 Hz, with
the snare usually containing more high-frequency energy than other mem-
branophones. Also, in the context of a given drum kit, the kick drum will
have a lower spectral centroid than the other membranophones. It can also
be noted that idiophones consisting of a metal plate will typically have their
spectral energy spread out more evenly across the frequency spectrum than
the membranophones, resulting in more high-frequency content.
Examples of three different drum instruments' time domain waveforms and
spectrograms are shown in Fig. 5.1. A kick drum is purely a membranophone,
containing a lot of low-frequency energy. A snare drum is also a membra-
nophone, but it has a snare belt attached below the lower membrane. When
the drum is hit, the lower membrane interacts with the snare belt, resulting in
a distinct sound also containing high-frequency energy. This can be observed
 5 Unpitched Percussion Transcription 133

in the figures. A crash cymbal is a plate-hke idiophone producing wide-band

energy. It is also worth noticing that the crash cymbal sound rings about ten
times longer than the example membranophone sounds.
The systems discussed in this chapter may process either signals that con-
sist of percussive instruments only, or complex music signals where the pres-
ence of pitched instruments is allowed. In many cases, the input to percussion
transcription systems will not contain recordings of live drums. Often the in-
put will consist of either sampled real drums, drum loops taken from another
recording, synthetic drums or even mixtures of the above. If a relatively small
set of drum samples are used, the resulting signal's waveform will exhibit less
variation between hits, making the signal easier to transcribe for many of the
presented systems. The same applies to the use of sampled drum loops and
synthetic drum sounds. Synthetically generated drum sounds tend to mimic
the overall timbral characteristics of real drums to some extent, but their spec-
tral characteristics tend to differ considerably from those of real drums. This
may cause problems if the transcription system is trained with real drums.
Approaches to percussion transcription can be roughly divided into two
categories: pattern recognition applied to sound events and separation-based
systems. The former segments the signal into meaningful events and recognizes
the contents of the segments with pattern recognition methods. This approach
is described in Section 5.2. The latter approach tries to separate the mixture
containing drum sounds in such a way that each segregated stream contains
only hits from one given drum type. Onsets are then sought from the streams.
Methods relying on this approach are described in Section 5.3. In addition,
a distinction can be made between systems that use a supervised approach,
through the use of trained classifiers or instrument templates, and systems
that use an unsupervised approach, such as clustering similar segments fol-
lowed by recognition of the clusters. Low-level signal analysis may not always
yield a satisfying result, so some attempts have been made to utilize musi-
cological modelling to take into account the predictability of drum patterns
in music. These are described in Section 5.4. Finally, some conclusions about
the presented methods are given in Section 5.5.

5.2 Pattern Recognition Approaches

In general, the percussion transcription problem can be characterized with two
questions: When did something happen in the music, and what was the event
that took place? The majority of the drum transcription systems developed
so far operate by answering these two questions in this same order. Here, such
systems are referred to as 'event-based' systems, and they tend to operate
with roughly the following steps:
1. Segment the input signal into events by
a) locating potential sound event onsets in the input signal, or
b) generating a regular temporal grid over the signal.
134 Derry FitzGerald and Jouni Paulus

2. Extract a set of features from each segment.

3. Classify the contents of each segment based on the extracted features.
4. Combine the segment time stamps with information about their content
to yield the transcription.
These steps will be described in more detail in the following sections.

5.2.1 Temporal Segmentation

There are two main approaches to segmenting a signal containing percussive

instruments. The first of these is based on detecting prominent sound event
onsets in an input signal, which is a difficult task in itself. The detection
method should be able to identify all the beginnings of meaningful sound
events and still be robust against noise which might generate extraneous on-
sets. This has led to the development of a number of different onset detection
methods which are detailed in Chapter 4 on p. 107.
An alternative to the onset detection-based segmentation is to generate
a temporal grid over the whole signal and to use it for the segmentation,
as suggested by Gouyon et al. [247]. The grid spacing is determined by the
fastest rhythmic pulse present in the signal, so that almost all the events in
the piece coincide with a grid point. A number of different names have been
given for this rhythmic pulse: Gouyon et al. called it the tick [247], Schloss, the
attack-point [567], and Bilmes, the tatum [37]. Here we use the term tatum.
There are several ways for estimating the tatum pulse, and they are dis-
cussed in more detail in Chapter 4. Gouyon et al. [247] and Seppanen [573]
detected onsets in the input signal, calculated time intervals between all onset
pairs, and determined the period of the tatum as the (approximate) greatest
common divisor of the time intervals. The phase of the grid was estimated
by aligning it with the located onsets. Another method of tatum estimation,
described by Klapuri et al., utilized a bank of comb filter resonators and
probabilistic modelling to find the tatum period and phase [349].
In practice, using a fixed equidistant grid has some problems. Even though
musical metre analysis procedures have developed considerably in recent years
and can often produce a reliable estimate of the tactus pulse (a.k.a the beat),
the tatum estimate tends to be unreliable. In particular, overestimating the
tatum period causes severe errors in the segmentation. Another drawback is
that expressive playing causes deviations from the equidistant grid points (see
for example [37] where Bilmes tried to construct a system capable of creating
similar expressive deviations to rhythmic patterns). An advantage of the grid
representation is that it is less prone to errors caused by inserting and deleting
sound onsets than the onset detection-based approach.
Before the subsequent feature extraction, meaningful parts of the signal
need to be segmented. The simplest approach is to take a part of the signal
starting at the located onset or grid point, and ending at the next located on-
set or grid point [245], [209]. However, if the consecutive onsets are far apart
 5 Unpitched Percussion Transcription 135

or do not have a constant time difference, it is good to limit the minimum and
maximum length of the segments. This guarantees that each of them contains
enough information for extracting relevant features. For example, good initial
guesses for the minimum and maximum lengths could be 50 ms and 200 ms,
respectively. A window function can be used in connection to this. However,
a traditional Hamming or Hanning window, for example, is not appropriate
since it smooths out the informative attack part at the beginning of the seg-
ment. A half-Hanning window which starts from a unity value and decays
to zero at the end of the segment is more suitable, but often windowing is
omitted completely, assuming that events decay to small amplitude naturally
and the signal does not contain sustained sounds at all.

5.2.2 Feature Extraction

The aim of feature extraction is to obtain numerical values describing the

segments so that they can be recognized or grouped together, while reducing
the amount irrelevant information in the time-domain signal. Quite a lot of
research has been carried out to find good descriptors for percussive instru-
ment classification [250], [245], [247], [287], [511]. Many of these features are
used also in pitched musical instrument recognition and are discussed in more
detail in Chapter 6. Here we will introduce only the features that are most
commonly used in percussion transcription. Ideally, the features should be
such that they are robust in the presence of other simultaneously occurring
sounds, but in practice this is difficult to achieve.
Mel-frequency cepstral coefficients (MFCCs) describe the rough shape of
the signal spectrum and are widely used in speech recognition [536]. Simi-
larly, they are often encountered in percussion transcription algorithms. A
detailed description can be found in Chapter 2, p. 25. Usually the coefficients
are calculated in short (about 20 ms) partially overlapping frames over the
analysed segment, and 5 to 15 coefficients are retained in each frame. Instead
of using these directly as features, typically the mean and variance of each
coefficient over the segment are used [211], [504], [209], [608]. In addition, the
first- and second-order temporal differences of the coefficients, and the means
and variances of these, are commonly used as features.
Another commonly used set of spectral features are bandwise energy de-
scriptors. The energy content of the sound is calculated in a few frequency
bands and their relations to the total signal energy are used as features. The
number of bands and their spacing depends greatly on the desired frequency
resolution, and systems have used from 6 [209] to 24 bands [285].
In addition to the rough spectrum described by the MFCCs and bandwise
energy descriptors, more simple spectral shape features are also useful. These
include the first four moments of the spectrum, i.e., spectral centroid, spec-
tral spread, spectral skewness, and spectral kurtosis [125], [307], [514]. Let us
denote the normalized magnitude spectrum by
136 Derry FitzGerald and Jouni Paulus

where X{k) denotes the discrete Fourier spectrum, /c is a frequency index,

and the set IC^ contains only non-negative frequency indices. The spectral
centroid is then defined as

The bandwidth of the spectrum is described by spectral spread,

Sf= Y.{k-C,fX{k). (5.3)

The spectral skewness describes the asymmetry of the frequency distribution

around the spectral centroid,

71 = ^3 (5-4)

Finally, spectral kurtosis describes the peakiness of the frequency distribution,

72 = ^^4 . (5.5)

The smaller the kurtosis, the flatter the spectrum. The quantities (5.2)-(5.5)
can also be calculated using a logarithmic frequency scale, as suggested in the
MPEG-7 standard [307].
In comparison with the spectral features, relatively few time-domain fea-
tures have been used in percussive sound classification. Instead, temporal
evolution of the sound is often modelled using differentials of spectral features
extracted in short frames over the segment. Among the features that can be
computed in the time domain, the two most commonly used are temporal
centroid and zero crossing rate. The temporal centroid, a direct analogue to
the spectral centroid, describes the temporal balancing point of the sound
event energy by

where E{t) denotes the root-mean-square (RMS) level of the signal in a frame
at time t, and the summation is done over a fixed-length segment starting
at the onset of the sound event. The feature enables discrimination between
short, transient-like sounds and longer ringing sounds. The zero crossing rate
describes how frequently the signal changes its sign. It correlates with the
spectral centroid and the perceived brightness of the signal. Usually, noise-
like sounds tend to have a larger zero crossing rate than more clearly pitched
or periodic sounds [250].
 5 Unpitched Percussion Transcription 137

Feature set is generally selected through trial and error, though some au-
tomatic feature selection algorithms have also been evaluated by Herrera et
al. [287]. It was noticed that in most cases, using a feature set that has been
chosen via some feature selection method yielded better results than using all
the available features. Also, a dimension reduction method such as principal
component analysis can be applied to the set of extracted features prior to
classification. For a more detailed description of feature selection methods and
possible transformations, refer to Chapter 2 and Chapter 6.

5.2.3 Segment Recognition

The extracted features are then used to recognize the percussive sounds in
each segment. There are at least two different ways to do this. The first is
to try to detect the presence of a given drum, even if other drums occur at
the same time, and the other is to attempt to recognize drum combinations
directly. For example, if an input signal consists of snare and hi-hat sounds,
the first approach will attempt to recognize the presence of both instruments
independently from each other, while the latter will attempt to recognize
whether 'snare', 'hi-hat', or 'snare + hi-hat' has occurred, treating sound
combinations as unitary entities.
A problem that arises when recognizing drum combinations instead of
individual drums is that the number of possible combinations can be very
large. Given M different drum types which may all occur independently, there
are 2^ possible combinations of them. That is, the number of combinations
increases rapidly as a function of M, and it becomes difficult to cover them all.
In practice, however, only a small subset of these combinations are found in
real signals. Figure 5.2 illustrates the relative occurrence frequencies of the ten
most common drum event combinations in a popular music database. These
contribute 95% of the drum sound events in the analysed data. When focusing
on the transcription of the drums commonly used in Western popular music,
the number of possible sound types M has usually been limited to the range
of two to eight. Some systems have concentrated on transcribing only the
kick and snare drum occurrences [235], [250], [221], [693], [608], [683], whereas
some others have extended the instrument set with hi-hats or cymbals [620],
[505], or added even further classes such as tom-toms and various percussion
instruments [506], [209].
Classification algorithms can be roughly divided into three different cate-
gories:
decision tree methods,
instance-based methods, and
statistical modelling methods.
With the exception of the work by Herrera et al. [287], there has not been an
extensive comparison of different classification methods as applied on percus-
sive sounds. Also, the experiments done in [287] concentrated on the classi-
138 Derry FitzGerald and Jouni Paulus

H BH HS S B BC BHS C T BCH
Instrument combination

Fig. 5.2. The relative frequencies of the ten most frequently occurring drum sound
combinations in the RWC Popular Music Database [229]. These combinations con-
tribute 95% of all drum sound events present in the database. A total of five drum
classes were used in the calculations, and are denoted as following: H is hi-hat, B is
kick drum, S is snare drum, C is cymbal, and T is tom-tom.

fication of isolated d r u m sounds instead of handling entire percussive sound

sequences. T h e results suggest t h a t the instance-based algorithms perform
best among the tested methods, and the decision tree algorithms the worst.
Here we will or^ly present briefly the methods t h a t have been used in unpitched
percussion transcription. For more detailed descriptions or other possible al-
gorithms, refer to Chapter 2 and Chapter 6.
Decision tree classifiers operate by asking a sequence of questions about
the sample to be classified. Such a question could be, for example, Ts the
spectral centroid of the signal above 500 Hz?' Each answer rules out some of
the possible classification results and defines the next question. Thus, the
classification process could be written out as a large conditional expression.
For a more detailed description, see [533], [161]. In addition to the recognition
of isolated d r u m sounds by Herrera et al. [287], Sandvold et al. [561] used
a decision tree classifier for the transcription of continuous percussive tracks
(see p. 142 below).
Instance-based methods store the given training samples (or a selected sub-
set of t h e m ) , and determine the label for the analysed sample by comparing
it to the stored training d a t a . In practical use, the principal weakness of tra-
ditional instance-based algorithms (e.g., /c-nearest neighbours) is the memory
required for storing the example events and the computational load needed
to compute distances to all the training samples. Support vector machines
(SVMs) remove these shortcomings by pre-processing the training d a t a so as
to retain only the samples t h a t actually have an effect on the classification.
T h e basic structure of a SVM is to locate a decision surface or a hyperplane
t h a t has the maximal margin with respect to the two classes it is trained
to separate [59], [568]. T h e decision surface is parameterized by the sample
p a t t e r n s t h a t have the smallest margin, called the support vectors.
SVMs have been used successfully in a number of classification tasks [568],
and also in percussion transcription [209], [608], [210]. Originally SVMs were
binary classifiers, so they have been used to detect whether or not t h e segment
contains a percussive sound event [506] and as a detector-like recognizer for
 5 Unpitched Percussion Transcription 139

individual drums [209], [608]. Subsequent extensions to enable multiple classes

have been developed; see [629]. Gillet and Richard evaluated the use of both
M binary SVMs and one 2 ^ class SVM in [209]. The results suggest that the
multiclass version performs slightly better than several binary classifiers. More
recently, the system was improved through decorrelating the used features
with principal component analysis, and adding implicit short-time context
modelling (see p. 155 for further details) in [210].
Gaussian mixture models (GMMs), a member of the statistical modelling
family, model the distribution of the feature values in each class as a sum
of Gaussian distributions. GMMs have proven to be a useful tool in pattern
recognition as they can be used to approximate arbitrary distributions. Fur-
ther details on GMMs can be found in Chapter 2 or in [536]. Paulus and
Klapuri used GMMs in percussion transcription by training a separate GMM
for each of the 127 non-silent combinations of different drum instruments [506].
More commonly, GMMs are used in conjunction with hidden Markov mod-
els (HMM) to represent the feature distributions in their states. A basic HMM
consists of two parallel processes: a hidden state process which is assumed
to be a Markov chain and cannot be directly observed, and an observation
process (features). The observed features are conditioned on the hidden state
by using a GMM in each state and, based on the observations, the hidden
state sequence can be inferred. Details on HMMs can be found in Chapter 2.
HMMs have been used in percussion transcription by Gillet and Richard,
who considered target signals consisting of Indian tablas'^ [211] or Western
drum sounds [209]. In [211] the different strokes were modelled by GMMs,
and HMMs were used to model event sequences in the tabla recordings. A
recognition rate of 94% was achieved on the test database. A similar approach
was used in [209], where GMMs in conjunction with HMMs were used to
transcribe drum loops. It was also noted that a SVM classifier without any
sequence modelling performed better than a HMM-based approach [209].
A recognition approach that cannot be put directly into any of the above
categories is the template matching method proposed by Goto et al. in
what was the first attempt at transcribing polyphonic drum mixtures [241],
[236]. The system aimed to identify mixtures of kick drums, snares, tom-
toms, hi-hats, and cymbals. The templates for each drum type were obtained
from an short-time Fourier transform-based power spectrogram of each drum
type in isolation. The templates were scaled to account for amplitude dif-
ferences between the templates and the mixture signal. A distance measure
was used to detect the presence of the template in the mixture. Detection of
membranophones was carried out using a logarithmic frequency resolution,
while idiophones were detected using a linear frequency resolution.
Later work by Goto identified snare and kick drum events in polyphonic
music as a means of tracking the beat in a piece of music [235]. This was done

^Tablas consist of a metallic bass drum and a wooden treble drum. Different
hand strokes on these drums produce different sounds.
140 Derry FitzGerald and Jouni Paulus

by first identifying potential onsets and then creating a histogram of spectral

peaks at these onset locations. The characteristic frequency of the kick drum
was identified as the lowest peak of the histogram, while the characteristic
frequency of the snare was identified as the largest peak in the histogram
above the characteristic frequency of the kick drum. It was then adjudged
that a kick drum or snare had sounded at times when an onset was detected
and the onset's peak frequency coincided with the characteristic frequency of
the drum in question. This method of detecting the snare was later replaced
by identifying noise components widely distributed in the frequency range of
1.4 kHz to 7.5 kHz as the snare drum. Unfortunately, no evaluation of the
system's performance in identifying the snare and kick drum was carried out
as it focused on beat tracking, not transcription.

5.2.4 Instrument Model Adaptation

Even though isolated percussive sounds can be identified quite reliably [285],
real-world recordings are not as easy to analyse. This is due to other simul-
taneously occurring interfering sounds, both other drums and melodic instru-
ments, as well as the fact that drum sounds can vary between occurrences,
depending on how and where they are struck. As a consequence, it is difficult
to construct general acoustic models that would be applicable to any data
and still discriminate reliably between different instruments.
A way to overcome this problem is to train the models with data that is as
similar as possible to the target mixture signals. However, this is not possible
if the exact properties of the target signals are not known in advance or they
vary within the material. Model adaptation has been proposed to alleviate
this problem. In this approach, the idea is to adapt general models to the
mixture signal at hands, instead of using fixed models for each and every
target signal. To date, only three event-based drum transcription systems
have been proposed that take this approach [693], [561], [683].
The earliest percussion transcription system utilizing model adaptation
was that of Zils et al. [693], which used an analysis-by-synthesis approach.
Initially, simple synthetic percussion sounds Zi{n) were generated from low-
pass and bandpass-filtered impulses. These represented very simple approxi-
mations to kick drums and snares respectively, and were then adapted to the
target signal to obtain more accurate models. The algorithm operated with
the following steps:
1. Calculate correlation function between a synthetic sound event Zi{n) and
the polyphonic input signal y{n)
Ni-l
^iW = X I ^iH2/(^ + r), (5.7)
n=0

where Ni is the number of samples in the sound i, and rj(r) is defined for
r [0, Ny Ni], where A^^^ is the number of samples in y{n).
 5 Unpitched Percussion Transcription 141

2. Locate occurrences of the sound Zi{n) in y{n) by picking peaks in ri{T)

and by retaining only the most rehable peaks.
3. Update the sound Zi{n) with

1 ^
Zi{n) ^ 2 :0,...,iV,-l, (5.8)
U
J

where U is the number of rehable peaks detected and TJ contains their

locations.
4. Repeat Steps 1 to 3 until convergence.
The above procedure was applied separately for the snare and kick drum.
When both drums occurred simultaneously, priority was given to the kick
drum, thus limiting the system to monophonic transcription of snare and kick
drum in the presence of pitched instruments. The system was tested on 100
examples from various music genres, and gave a success rate of over 75% in
cases where the percussion instruments were louder or as loud as the other
instruments in the mixture. In cases where the percussion instruments were
quieter, a 40% success rate was reported.
The system described by Yoshii et al. transcribed snare and kick drums
in the presence of pitched instruments by using an analysis-by-synthesis ap-
proach in the time-frequency domain [683]. First, onset detection was carried
out and excerpts Pj{t^k) of the mixture power spectrogram were extracted
from the positions of the onsets. These were then used to adapt snare and kick
drum templates (models) which were initialized using the spectrograms of iso-
lated examples. The following template-adapt at ion algorithm was repeated for
both target drum classes:
1. Calculate the Euclidean distances Fij between a template Ti and the
spectrogram excerpts Pj with
T-lK-l
^h = E E [HMk){Ti{t,k)-P,{t,k))]\ (5-9)

where T is the number of frames in the template and spectrogram excerpt,

K is the number of frequency bins, and the response Hi,p{k) is used to
attenuate high frequencies.
2. Order the spectrogram excerpts by their distances Fij in ascending order.
3. Choose a fixed fraction (10%) of the excerpts with the smallest distances,
and calculate a new template spectrogram with
Ti{t, k) <~ medid^njes{Pj{t^ A:)}, (5.10)
where the set S contains the chosen excerpts.
4. Repeat Steps 1 to 3 until convergence.
The median was used in an attempt to suppress the effect of the presence of
pitched instruments on the drum templates.
142 Derry FitzGerald and Jouni Paulus

The actual segment recognition was done with a template-matching method

stemming from the work of Goto and Muraoka [236] (see p. 139). The system
was tested on 10 songs from the RWC database [229], and gave a recall rate^
of 90% and a precision rate of 90% for the kick drum, and recall and precision
rates of 83% and 93% for the snare drum. The use of adaption improved the
results significantly over simple template matching.
In the system described by Sandvold et al. [561], the adaption took place
at a higher abstraction level than in the previous systems. Here the models
used were based on extracted features, rather than a time domain or time-
frequency domain template. After onset detection, segmentation and feature
extraction, the algorithm consisted of the following steps:
1. Classify the extracted n segments with a decision tree classifier and general
instrument models (see Section 5.2.3).
2. Evaluate the reliability of the classification results, and choose m < n
most reliable results.
3. Create signal-specific (localized) models using the chosen m events.
4. Classify the contents of all the n segments using the localized models.
The general instrument models were constructed by extracting 115 spectral
and temporal features from the training data and by choosing the most suit-
able subset of them with a correlation-based feature selection algorithm (see
Chapter 6, p. 182). The reduced feature set consisted on average of less than
25 features for each class. When constructing the localized models in Step 3,
all the 115 features were reconsidered and the best subset was again chosen
with the same method. In this case, nine features per class were sufficient on
the average. The ranking and selection of the most reliably classified events
in Step 2 was not done fully automatically. Instead, the result from the clas-
sification with general models was corrected and the reliability ranking was
done manually, thus making the overall system operate semi-automatically.
The classification algorithm used with the localized models was the A:-nearest
neighbours algorithm with fc = 1, i.e., a simple instance-based algorithm.
The method was evaluated with seventeen 20-s excerpts from polyphonic
audio recordings with manually annotated ground truth. The results suggest
that the use of localized models reduces the required number of features (the
average feature set size drops from 25 to 9), and improves the recognition
result, with the average accuracy improving from 72% to 92%.

5.3 Separation-Based Approaches

Having explored event-based approaches for percussion transcription, an al-
ternative way of approaching the problem is through the use of source
^Recall rate is defined as the number of correctly transcribed events divided by
the number of events in the input. Precision rate is defined as the ratio of correctly
transcribed events to the total number of events at the output of the system.
 5 Unpitched Percussion Transcription 143

separation-based techniques. These have some advantages over the event-

based techniques, while at the same time having drawbacks of their own.
Separation-based techniques can facihtate the analysis of mixtures of drums
or percussion instruments by separating the instruments into distinct streams,
which makes the task of transcribing easier. However, a drawback of many
separation-based methods is that the streams have to be identified after sep-
aration. Further, these techniques often have difl^iculties in detecting low-
intensity sounds. Despite these limitations, separation-based methods have
proven to be a viable way of approaching the problem of percussion
transcription.
The most commonly used source separation technique is independent com-
ponent analysis (ICA) [96]. Further details on ICA can be found in Chapter 9.
However, the basic ICA requires as many sensors (microphones) as there are
sources for separation to occur, whereas in the drum transcription task there
are usually at most two channels available; in other words, a stereo signal.
Further, some drums will often be of the same amplitude in both channels,
particularly snare and kick drums, and so in effect there is only a single channel
available to separate these drums. As a result, most work using separation-
based approaches has focused on single-channel separation.
Various methods for separating sound sources from single-channel record-
ings have been used, such as independent subspace analysis (ISA) [73], non-
negative sparse coding (NNSC) [299], and non-negative matrix factorization
(NMF) [400], and all have found uses in the area of percussion transcription.
ISA has also been used for rhythmic analysis [488]. A detailed description of
these methods can be found in Chapter 9. Here, discussion of these source
separation techniques is limited to how they relate to the problem of per-
cussion transcription. It is important to note that separation in the context
of transcription means the separation of frequency and amplitude characteri-
stics associated with each source in order to identify and transcribe them,
as opposed to resynthesis of the separated sources, though resynthesis of the
separated sources is also possible using these techniques. Methods of source
resynthesis are presented by FitzGerald in [186].
All these techniques assume that the mixture spectrogram matrix X of
size {K X T), where K is the number of frequency bins and T is the number
of time frames, results from the superposition of J source spectrograms Y^
of the same size. Further, it is assumed that each of the spectrograms Yj
can be uniquely represented by the outer product of an invariant frequency
basis function hj of length K and a corresponding invariant amplitude basis
function (or time-varying gain) gj of length T which describes the gain of the
frequency basis function over time. This yields
J J

x = E Y ; = Eb^gJ- (5.11)
3=1 i=l
144 Derry FitzGerald and Jouni Paulus

The above decomposition is typically applied to a magnitude spectrogram,

though power spectrograms can also be used.^ Where the techniques differ is
in how the decomposition of the spectrogram X into frequency and ampli-
tude basis functions is achieved. The basic ISA method proposed by Casey
did this by performing principal component analysis (PCA) [320] on the spec-
trogram, keeping only a small number of decorrelated frequency lines, and
then performing ICA on the retained components [73]. Further details on
PCA can be found in Chapter 2, while details on ISA and ICA can be found
in Chapter 9. In effect, ISA performs ICA on a low-dimensional representation
of the original spectrogram. NNSC attempts to balance modelling accuracy
with the sparseness of the recovered sources while enforcing non-negativity
on the sources [299], whereas NMF attempts to reconstruct the data using
non-negative basis functions and a Poisson or Gaussian noise model [400].
In practice, the use of invariant frequency basis functions means that no
pitch changes are allowed over the course of individual spectrograms Y j . How-
ever, this is valid for most drum sounds, where the pitch of the drum does not
change from event to event, making this type of decomposition particularly
suited for analysing percussive tracks in polyphonic music.
Figure 5.3 shows the magnitude spectrogram of a drum loop containing
snare and kick drum hits, while Fig. 5.4 shows the associated amplitude and
frequency basis functions recovered using NMF. As can be seen from the am-
plitude basis functions, the amplitude envelopes of both the snare and the kick
drum have been well separated, though some evidence of the kick drum is still
visible in the snare basis function. The frequency basis functions can be seen
to have captured the overall spectral characteristics of the sources, with the
kick drum having more low-frequency energy than the snare, and the snare
having its energy spread over a wider frequency range. This demonstrates
the usefulness of these techniques for the purposes of drum transcription.
Nonetheless, from the point of view of percussion transcription, there are
certain problems with the above-mentioned techniques, regardless of how the
decomposition is achieved. These can be summarized as follows.
Indeterminate source order: sources have to be identified after separation.
Estimation of the optimal number of basis functions: keeping a small num-
ber of basis functions results in more recognizable features, while recover-
ing low-energy sources requires increased numbers of basis functions.
Further, ISA recovers basis functions which may have negative elements,
which does not reflect the assumption that the overall magnitude or power
spectrogram results from the summation of independent spectrograms which
are non-negative by definition. This can lead to errors in transcription. How-
ever, this is not a problem for NNSC and NMF, which constrain non-negativity
of the sources. FitzGerald discusses these problems in greater detail in [186].
^The term 'magnitude spectrogram' refers to a representation which consists of
the absolute values of the discrete Fourier transform in successive time frames. A
power spectrogram is obtained as the element-wise square of this.
 5 Unpitched Percussion Transcription 145
rf " 1 1 ' i

> 1
1 1

1 1

1.5 2
Times

Fig. 5.3. Magnitude spectrogram of a drum loop containing snare and kick drum.

A number of techniques have been developed in an attempt to overcome the

above problems, all using prior knowledge about the drums to be transcribed
[187], [188], [505]. The first of these added sub-band pre-processing to ISA
[187]. As noted in Section 5.1, the drums in a standard rock kit can be divided
into membranophones and idiophones. Membranophones have most of their
energy in the low end of the frequency range, while metal plate idiophones
have most of their energy spread over the spectrum above 2 kHz. Filtering the
signal using a lowpass filter with a 1-kHz cut-off frequency and a highpass filter
with a 2-kHz cut-off frequency allows the membranophones and idiophones to
be emphasized, respectively.
This was demonstrated in a system described in [187] that transcribed
mixtures of snare, kick drum, and hi-hats. Two basis functions were found to
be sufficient to recover the snare and kick drum from the lowpass-filtered signal
using ISA, and similarly two basis functions recovered the hi-hats and snare
from the highpass-filtered signal. To overcome the problem of source ordering,
it was assumed that the kick drum had a lower spectral centroid than the snare
drum, and that hi-hats occurred more frequently than the snare drum.
The system was tested on a set of 15 drum loops consisting of snare,
kick drum, and hi-hats. The drums were taken from a number of different
drum sample CDs and were chosen to cover the wide variations in sound
within each type of drum. A range of different tempos and metres were used
and the relative amplitudes between the drums varied between 0 dB and
24 dB. In total, the test set contained 133 drum events. The success rate was
evaluated using c = {t u i)/t where c is the percentage correct, t is the total
number of drums, u is the number of undetected drums, and i is the number of
146 Deny FitzGerald and Jouni Paulus
2.51-
2\
J , , , , ' -
1.5h
-
-
0.51
Iv 0.5 1
V , JV_:

100 200 300 400 500 600 700 800 900 1000
Frequency Hz

Fig. 5.4. Basis functions recovered from the spectrogram in Fig. 5.3. Prom top
to bottom, they are the kick drum ampUtude basis function, the snare drum am-
phtude basis function, the kick drum frequency basis function, and the snare drum
frequency basis function.

incorrectly detected drums. Using this measure, an overall success rate of 90%
was achieved. However, more effective means of incorporating prior knowledge
were subsequently developed, and are discussed in the following subsections.

5.3.1 Prior Subspace Analysis

Prior subspace analysis was proposed as a means of tackling the problem of

percussion transcription by combining prior models of drum sounds with the
source separation techniques mentioned above [188]. PSA uses the same signal
model as shown in (5.11) earlier. PSA then assumes that there exists known
frequency subspaces or basis functions bprj that are good initial approxima-
tions to the actual subspaces (here the terms subspace and basis function are
used as synonyms). In other words, it is assumed that frequency basis func-
tions such as shown in Fig. 5.4 are available before any analysis of the signal
takes place, and that these are good approximations to the actual frequency
basis functions that could be recovered from the signal. Substituting bj with
these prior subspaces yields

prj^j (5.12)
 5 Unpitched Percussion Transcription 147

In matrix notation, this becomes

X^BprG, (5.13)

where Bpr = [bpri,. , bpr j] and G = [ g i , . . . , gj] . Estimates of the am-

phtude basis functions can be obtained from

G = Bpr+X, (5.14)

where ^ denotes the pseudoinverse. The pseudoinverse of a matrix B can

here be computed as B+ = (B^B)~^B^, assuming that bprj are hnearly
independent. The estimated amphtude basis functions gj are not independent,
and may contain information related to more than one source. This is as
a result of the broad-band noise-based nature of drum sounds, where the
occurrence of a given drum will cause a partial match with the prior subspace
for another drum.
To overcome this problem and to recover amplitude basis functions which
are related to only a single source, ICA is carried out on the estimated am-
plitude basis functions G to give

G-WG, (5.15)

where W is the unmixing matrix obtained from ICA and G contains the inde-
pendent amplitude basis functions. This results in amplitude basis functions
which are generally associated with a single source, though there will still be
some small traces of the other sources. Improved estimates of the frequency
basis functions can then be obtained from

B = XG+. (5.16)

In this case, the use of the pseudoinverse is justified in that the columns of G"^
are orthogonal and do not share any information, and the pseudoinverse can
be calculated as G"^ = G^(GG^)~^. The overall procedure can be viewed as
a form of model adaptation such as is described in Section 5.2.4.
Figure 5.5 shows a set of priors for snare, kick drum, and hi-hat, respec-
tively. These priors were obtained by performing ISA on a large number of
isolated samples of each drum type and retaining the first frequency basis
function from each sample. The priors shown then represent the average of
all the frequency basis functions obtained for a given drum type. Priors could
be obtained in a similar way using some other matrix-factorization technique
such as NMF (see Chapter 9 for further details). It can be seen that the priors
for both kick drum and snare have most of their energy in the lower regions
of the spectrum, though the snare does contain more high-frequency informa-
tion, which is consistent with the properties of membranophones, while the
hi-hat has its frequency content spread out over a wide range of the spectrum.
The use of prior subspaces offers several advantages for percussion in-
struments. First, the number of basis functions is now set to the number of
148 Derry FitzGerald and Jouni Paulus

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Frequency kHz

1 AMi \k
0.8
0.6
0.4
i'VVvw^/ kP ^
\u.
n^i 'yf(

^L^f^^
0.2 MM\J^
"""^"'^^
' " ^ " ^ .

0 5 10 15 20
Frequency kHz

Fig. 5.5. Prior subspaces for snare, kick drum, and hi-hat.

prior subspaces used. Second, the use of prior subspaces alleviates the bias
towards sounds of high energy inherent in blind decomposition methods, al-
lowing the recovery of lower-energy sources such as hi-hats. Thus, the use of
prior subspaces can be seen to go some distance towards overcoming some of
the problems associated with the use of blind separation techniques, and so
is more suitable for the purposes of percussion transcription.
A drum transcription system using PSA was described in [188]. Again,
the system only transcribed signals containing snare, kick drum, and hi-hats
without the presence of any other instruments. Prior subspaces were generated
for each of the three drum types, and PSA performed on the input signals.
Once good estimates of the amplitude basis functions had been recovered,
onset detection was carried out on these envelopes to determine when each
drum type was played. To overcome the source-ordering problem inherent in
the use of ICA, it was again assumed that the kick drum had a lower spectral
centroid than the snare, and that hi-hats occurred more frequently than the
snare. When tested on the same material as used with sub-band ISA (see
p. 145 for details), a success rate of 93% was achieved.

5.3.2 Non-Negative Matrix Factorization-Based Prior Subspace

Analysis

More recently, an improved formulation of PSA has been proposed for the
purposes of drum transcription [505], based on using an NMF algorithm with
 5 Unpitched Percussion Transcription 149

priorly fixed frequency basis functions Bpr- The NMF algorithm estimates
non-negative amphtude basis functions G so that the reconstruction error of
the model (5.13) is minimized.
This offers a number of advantages over the original formulation of PSA.
First, the non-negative nature of NMF is more in keeping with the data being
analysed, in that the spectrogram is non-negative, and so a decomposition
that reflects this is likely to give more realistic results. Second, keeping Bpr
fixed eliminates the permutation ambiguities inherent in the original PSA
algorithm. This allows the elimination of the assumptions necessary to identify
the sources after separation, and permits the algorithm to function in a wider
range of circumstances.
The NMF-based algorithm estimates G by first initializing all its elements
to a unity value and then iteratively updating the matrix using the rule

G.x B p / ( x . / ( B p r G ) ) ] ./ [ B p / l ] , (5.17)

where .x and ./ denote elementwise multiplication and division, respectively,

X is the signal spectrogram, and 1 is an all-ones matrix of size equal to X.
When tested on the same test material as subband ISA (see p. 145), a success
rate of 94% was obtained.
A transcription system using the NMF-based algorithm was presented by
Paulus and Virtanen [505]. The mid-level signal representation was a mag-
nitude spectrogram with only five frequency bands (20-180 Hz, 180-400 Hz,
400-1000 Hz, 1-10 kHz, and 10-20 kHz), with an analysis frame length of
24 ms and frame overlap of 75%. The fixed instrument priors were calculated
from a large set of training samples, by factorizing each sample with NMF
into an outer product of an amplitude basis function and a frequency basis
function. The frequency basis functions of all samples of a given individual
sound type were averaged to yield bpr. The time-varying gains were calcu-
lated by iteratively applying (5.17) while the source spectra were kept fixed.
The onsets were detected from the calculated gains with a variation of the
onset detection algorithm described in [349].
The performance of the transcription system was evaluated with acoustic
material consisting of different drum kits recorded in different environments,
in total giving four different recording sets. In each recording set, five different
sequences were recorded, resulting in a total of 20 signals. The sequences were
fairly simple patterns containing mostly kick drum, snare drum, and hi-hat.
Two monophonic mix-downs of the signal were produced: a dry, unprocessed
mix with only levels adjusted, and a wet, 'production grade' mix with com-
pression, reverb and other effects, which attempted to resemble drum tracks
on commercial recordings. Also evaluated using this material were the SVM
classifier-based event recognition method [209], as implemented by Paulus,
and the earlier PSA method [188] as implemented by the original author.
All three systems were trained with the same material to the extent that
was possible. The spectral basis functions for the NMF system and PSA were
150 Derry FitzGerald and Jouni Paulus

calculated from the same set of unprocessed samples. For the NMF-based
system, onset-detection thresholds were obtained by analysing a set of training
signals and setting the threshold to a value minimizing the number of detection
errors. For PSA, the source-labelling rules and fixed threshold values from the
original publication were used. The SVM-based classifier was trained so that
when analysing the dry (wet) signals, the training features were also extracted
from the dry (wet) signals. This may have given the SVM method a shght
advantage compared to the other systems.
The NMF-based system performed best of the methods, with the dry mix
material having a hit rate of 96% compared to the 87% of the SVM method
and 67% of PSA. The performance gap became smaller with the production-
grade mixes, but the NMF-based method still had a hit rate of 94% compared
to the 92% of SVM method and 63% of PSA [505].

5.3.3 Prior Subspace Analysis in the Presence of Pitched

Instruments

A modified version of the original PSA algorithm was found to be effective

in transcribing snare, kick drum, and hi-hats/ride cymbals from recordings of
polyphonic music [189]. The modifications were required because the presence
of a large number of pitched instruments will cause a partial match with the
prior subspace used to identify a given drum, which can make detection of
the drums more difficult. However, it should be noted that pitched instruments
have harmonic spectra with regions of low intensity between the overtones or
partials. As a result, when pitched instruments are playing there will be re-
gions in the spectrum where little or no energy is present due to pitched
instruments. It was observed that good frequency resolution reduced the in-
terference due to the pitched instruments, and so improved the likelihood of
recognition of the drums.
In the case of snare and kick drum, setting all values in the initial estimates
of the amplitude basis functions G below a fixed threshold to zero was found
to be sufficient to eliminate the interference. ICA was then performed on these
thresholded basis functions to obtain better estimates for snare and kick drum.
In the case of hi-hats/ride cymbals, the interference was often considerably
worse, and simple thresholding was insufficient. However, most of the energy in
pop and rock sounds is in the lower region of the spectrum, and, as already
noted, the energy of the hi-hats is spread across the spectrum. Therefore, the
average power spectral density (PSD) [639] over the whole duration of each
excerpt was calculated, and the spectrogram then normalized by it. This is
equivalent to highpass filtering the signal, but in each case the filter takes
into account the characteristics of the signal being filtered. The hi-hat prior
is then multiplied with the PSD-normalized spectrogram to obtain the hi-hat
subspace.
The modified algorithm then consisted of carrying out ICA on the thresh-
olded snare and kick drum subspaces, followed by onset detection to identify
 5 Unpitched Percussion Transcription 151

these drums, and then onset detection on the hi-hat subspace recovered from
the PSD-normahzed spectrogram. As the hi-hat subspace no longer under-
goes ICA with the other drums, the algorithm loses the ability to distinguish
between a snare on its own and a snare and hi-hat occurring simultaneously.
Fortunately, in many cases these drums do occur simultaneously and so this
results in only a small reduction in the efficiency of the algorithm. When tested
on a database of 20 excerpts from pop and rock songs taken from commercial
CDs, an overall success rate of 83% was achieved.
Attempts to extend the basic PSA method to include other drums such
as tom-toms and cymbals met with mixed success. Extensive testing with
synthetic signals revealed that this was due to the fact that when the main
regions of energy of different sources overlap, as is often the case with drums
such as snares and tom-toms, then the sources will not be separated correctly
[186].

5.3.4 Input-Generated Priors

Attempts have been made to overcome the limitations of PSA by generating
frequency basis functions from the input signal which can then be used as
priors in a PS A-type framework, rather than using general priors. These basis
functions should then be better able to exploit differences in the spectra of the
drums present. To date, two systems have made use of this approach [190],
[147].
The system described by FitzGerald in [190], which worked on signals con-
taining drums only, models each event that occurs in the signal and generates
similarity measures between each event, which are used to group the events.
These groups are then used to generate frequency basis functions from the
input signal which can then be used as priors.
Membranophone events were identified by onset detection in the amplitude
envelopes obtained from
G = Bp/X, (5.18)
where, in this case, Bpr contains only snare and kick drum priors. This was
also sufficient to recover tom-tom onsets. Onset times were then determined,
and the sections of the spectrogram between each event obtained. PCA was
then used to obtain a frequency basis function to represent each section. The
resulting frequency basis functions were then clustered and an average fre-
quency basis function was obtained for each of the groups.
To overcome the most commonly occurring membranophone overlap, that
of snare and kick drum, the groups most likely to correspond to snare and
kick drum were identified. All tom-tom events in the original spectrogram
were then masked. PSA was performed on the resulting spectrogram, and the
snare and kick drum events identified. Other membranophone overlaps were
assumed not to occur as they are not very common, as can be seen in Fig. 5.2.
To detect idiophones, (5.18) is applied to a PSD-normalized spectrogram
using only a hi-hat prior subspace to yield an idiophone amplitude envelope.
152 Derry FitzGerald and Jouni Paulus

As traces of both snare and tom-toms will occur in the idiophone envelope,
an amplitude envelope for snare/tom-toms is obtained by masking kick drum
events in the original spectrogram, and multiplying the resulting spectrogram
by a snare frequency subspace, again using (5.18). ICA is then performed to
separate the snare/tom-tom amplitude envelope and the idiophone amplitude
envelope. Onset detection on the resulting independent idiophone envelope
then yields the idiophone events.
Grouping is then carried out on the idiophones. If two large groups occured
that did not overlap in time, then both hi-hat and ride cymbal were assumed
to be present; otherwise all events were allocated to the same drum. The
justification for this is detailed in [190]. Unfortunately, though the algorithm
distinguished between ride cymbal and hi-hats, it did not identify which was
which. When tested on a database of 25 drum loops, a success rate of 90%
was obtained using the same measure as sub-band ISA.
Dittmar et al. described a system which attempted to transcribe drums in
the presence of pitched instruments [147]. To enable recovery of low-energy
sources such as hi-hats and ride cymbals, the high-frequency content of the
signal was boosted in energy. A magnitude spectrogram of the processed sig-
nal was obtained, and then differentiated in time. This suppressed some of
the effects of the sustained pitched instruments present in the signal, be-
cause their amplitudes are more constant on a frame-by-frame basis than
that of transient noise, and so when differentiated will have a smaller rate of
change.
Onset detection was then carried out and the frame of the difference spec-
trogram at each onset time extracted. As the extracted frames contain many
repeated drum events, PCA was used to create a low-dimensional represen-
tation of the events. J frequency components were retained and non-negative
ICA [526] performed on these components to yield B, a set of independent
basis functions which characterized the percussion sources present in the
signal. The amplitude envelopes associated with the sources were obtained
from
G - B'^X, (5.19)
where G are the recovered amplitude envelopes, and X is the original spec-
trogram. A set of differentiated amplitude envelopes was then recovered
from
G - B'^X', (5.20)

where G are the differentiated amplitude envelopes and X' is the differentiated
spectrogram. Correlation between G and G was used to eliminate recovered
sources associated with harmonic sounds, as sustained harmonic sources will
tend to have lower correlation than percussive sources.
 5 Unpitched Percussion Transcription 153

The problem of determining the required number of components was ap-

proached by retaining a set number of components and then matching the
recovered components to percussion instrument templates. Drum onsets were
then identified by onset detection on the recovered amphtude envelopes.

5.3.5 Non-Negative Sparse Coding

The system proposed by Virtanen made use of NNSC to transcribe drums
in the presence of pitched instruments [650]. The NNSC algorithm used was
similar to that described by Hoyer in [299], with additional constraints to
encourage temporal continuity of the sources. The optimization was carried
out to obtain sparse amplitude basis functions, but sparse spectra were not
assumed. A power spectrogram was used as the input to the system.
The described transcription system attempted to transcribe snare and kick
drums in the presence of pitched instruments, and was tested using signals
synthesized from MIDI files. The system separated a pre-defined number of
basis functions, and then searched for frequency basis functions that matched
well with the template spectra of either the snare or kick drum. The goodness
of fit of a given source to a given template was calculated as

' [b^]fc-f6
(5.21)
k=l I
[^pr i]k + e

where hpri denotes the template spectrum of drum z, bj is the frequency

basis function for source j , /c is the frequency bin index, K is the number of
frequency bins, and e is a small positive value to ensure that the logarithm is
robust for small values.
Once sources associated with both snare and kick drum had been identi-
fied, onset detection was done on the amplitude envelopes of these sources.
An overall success rate of 66% was obtained for this system. The templates
were obtained by separating a number of sources in training material, and
averaging the frequency basis functions of the sources whose amplitude ba-
sis functions showed temporal correlation with the kick or snare drums in a
reference annotation.
As can be seen from the above, the use of source separation methods has
shown great utility for the purposes of unpitched percussion transcription,
overcoming some of the disadvantages of event-based systems, but at the cost
of some other problems. Further, despite the success of such systems to date,
there still remains much work to be done in improving the performance of
separation-based percussion transcription systems.

5.4 Musicological Modelling

Many of the presented transcription systems concentrate only on low-level
recognition, without any processing at a higher abstraction level. When
154 Derry FitzGerald and Jouni Paulus

compared to methods for automatic speech recognition, this resembles the

situation where each phoneme is recognized independently of its context. It
was noted quite early that by utilizing linguistic knowledge, speech recognition
performance could be improved considerably [321]. It is reasonable to assume
that incorporating similar ideas in music transcription can result in improved
performance. Here we use the term 'musicological modelling' to refer to at-
tempts to model the statistical dependencies of event sequences. There has
been relatively little work done utilizing such models in percussion transcrip-
tion. However, some methods have been proposed, and these are discussed in
the following [590], [506], [504], [211], [209].

5.4.1 Enabling High-Level Processing

Usually, high-level processing consists of some sort of probabilistic modelling

of the temporal relations between sound events. In practice, this requires that
the low-level recognition results must also be presented as a probability or
likelihood value. Some classification methods, for example Gaussian mixture
models, produce probabilities as their output directly, whereas the result of the
other methods has to be mapped to a probability value. There has not been
much research on how to do this with different classifiers, with the exception
of the work done on moderating SVM outputs into probability values [381],
[524], [423], [424], [156]. The different methods have been evaluated by Riiping
[556], and the results suggest that Piatt's method [524] performs best.
The output of a binary SVM classifier is a real number, where the decision
border between the classes resides at the value zero. Usually, the classification
is done according to the sign of the output value. Given the SVM output value
fi of the zth sample, Piatt proposed that the probability of the sample to be
from the positive class can be calculated with a sigmoid

where a and /? are sigmoid parameters estimated by Piatt's algorithm.

Another possible way of estimating likelihood values from the numerical
result of a non-probabilistic classifier is to utilize a histogramming method
called binning, described by Zadrozny and Elkan [686]. The output values fi
of the classifier are analysed for the training data, and a histogram of B bins
is constructed. At the recognition stage, each output value of the classifier
can be assigned to one of the bins. The probability of the classified sample to
belong to class c is then calculated as the fraction of the training samples in
the assigned bin that actually were from the class c.

5.4.2 Short-Term Context Modelling

Contextual information can be represented with a model separate from the

low-level signal analysis, or it can be taken into account within the low-level
 5 Unpitched Percussion Transcription 155

recognition algorithm. In the former, the context modelhng is done exphc-

itly, whereas in the latter, it takes place more implicitly. Several methods for
modelling sequential dependencies are reviewed by Dietterich [146]. Among
the implicit methods, only the sliding window technique has been used in
percussion transcription [210].
The sliding window method utilizes the feature values of the neighbouring
segments in the recognition of a given segment. In addition to the feature
vector Oi of the segment z, the features Oi-^^^..., o^,..., O^-^A^^ of the sur-
rounding 2A^w segments are used by concatenating them into one large feature
vector which is used for the classification. The window of features is slid over
the whole signal to be analysed. The sliding window method was utilized by
Gillet and Richard in [210] to improve the performance of their earlier method
[209]. The applied window was very short, utilizing only the features of one
preceding event. Combined with decorrelating the features with PC A, the
gained performance improvement was on average 3 to 4 percentage units.
The simplest form of explicit musicological modelling is the use of prior
probabilities for events. Depending on the selected approach, this means either
the probabilities of individual drum events or different drum combinations to
be present. Paulus and Klapuri applied the latter type of priors in a system
which was trained to recognize 127 different drum combinations (see p. 139
above) [506]. By using prior probabilities that resembled the profile in Fig. 5.2,
the error rate of the system decreased from 76% to 50%.
Even though the prior probabilities themselves are helpful in the over-
all task, they cannot model the sequential dependencies between consecutive
events. In percussive sequences, as with natural languages, the preceding con-
text makes some continuations more probable than the others. This sequential
dependence can be modelled with A^-grams.
The AT-grams rely on the Markov assumption that the event Wk, occur-
ring at the time instant k, depends only on the preceding N 1 events
Wk-N+ii"' I'^k-i^ instead of the complete history. Using wi:k to denote a
sequence of events wi,W2^. - ,Wk^ the Markov assumption can be written as

p{Wk\wi:k-l) = p{Wk\Wk-N-^l:k-l)' (5.23)

To enable the use of A-grams, the concept of an 'event' itself has to be

defined. For example, Paulus and Klapuri [506] estimated a grid of equidistant
tatum pulses over the performance and considered the combination of drum
sounds starting at grid point A: as a mixture-event Wk, for which A'-gram
models were estimated and used.
Usually the A'-gram probabilities are estimated from a large body of train-
ing material by
/ I X C{Wk-N-^l:k) /p, r>.x
p[Wk\Wk-N-\-l:k-l) = -pTi ^' (^24)
(^{Wk-N+l:k-l)

where C{wk-N+i\k) is the number of occurrences (count) of the sequence

Wk-N-\-i:k in the training material [321]. The set of probabilities for all possible
156 Derry FitzGerald and Jouni Paulus

sequences of length N is called the A^-gram model. These can be used to assess
the likelihoods of different event sequences in percussion transcription, or to
predict the next event.
If the events in the sequence are drawn from a dictionary of size D, there
are D^ probabilities that need to be estimated for an TV-gram of length N.
This imposes requirements on the size of the training data set in order that the
resulting A/'-grams do not contain too many zero-probability entries. Usually,
such entries cannot be completely avoided, so methods for reducing their effect
have been developed. The zero-probabilities can either be smoothed (given a
non-zero value) with a discounting method like Witten-Bell discounting [673]
or Good-Turing discounting [218], [90], or the required probability can be
estimated from lower-order A'-grams with the back-off method suggested by
Katz [336] or with the deleted interpolation algorithm by Jelinek and Mercer
[317]. An interested reader is referred to the cited publications for details.
The sound event A"-grams in music analysis are directly analogous to the
word A^-grams in speech recognition. Moreover, as the words in speech are con-
structed from individual letters or phonemes, the mixture-events in percussive
tracks may consist of multiple concurrent sounds from different instruments.
The main difference between these two is that in speech recognition, the order
of individual letters is important and the letters in consecutive words rarely
have any direct dependence, whereas in musicological A'-grams, the mixture-
events consist of co-occurring sound events which alone exhibit dependencies
between the same sound event in the neighbouring mixtures. This observa-
tion can be utilized to construct AT-gram models for individual instruments,
as suggested by Paulus and Klapuri [506].
When the set of possible instruments {ui,U2,..., UM} is defined, with the
restriction that each instrument can only occur once in each mixture-event,
the problem of estimating mixture-event A^-grams can be converted into the
problem of estimating A'-grams for individual instruments. In other words,
at each time instant k the instrument Ui has the possibility of being present
in the mixture Wk or not. By using this assumption, the probability estimate
from (5.23) becomes

p{Wk\Wk-N+l:k-l) = Yl Pi^iM'^hk-N-^l-.k-l)

Uj ^Wk

where p{ui^k) is the probability of the instrument Ui to be present at the

temporal location k.
Separate A'-grams for individual instruments have a clear advantage over
the A'-grams for mixture-events when considering the training of the model.
If there are M different instruments available, then there are 2 ^ different
mixture-events, requiring a total of 2^^^ probabilities to be estimated for the
mixture-event A^-grams. When using separate A'-grams for the instruments.
 5 Unpitched Percussion Transcription 157

each model has only a binary dictionary, leading to a total of M2^ probabili-
ties to be estimated. This alleviates the zero-frequency problem significantly:
the sharply concentrated prior distribution of different mixture-events (see
Fig. 5.2) means that some of them occur too rarely for reliable probability
estimation, even in a large training set.
The main problem with individual instrument A^-grams is that they lack
information about simultaneously occurring events. Each A^-gram 'observes'
only the presence of its own instrument without any knowledge of the other
co-occurring instruments. As a result, the model may give overly optimistic
or pessimistic probabilities to different mixture-events. One possible way to
address this problem is to use the prior probabilities of the mixture-events in
connection with (5.25), as proposed in [506].
Musicological prediction can also be done using simpler modelling. For
example, if two occurrences of the same event type took place with time
interval t/^, its occurrence can be predicted again after another interval of IA-
A system relying on this type of modelling was proposed by Sillanpaa et al.
in [590].

5.4.3 Modelling of Periodic Dependencies

The fact that A^-grams only use the directly preceding events to predict the
next event is a minor drawback, considering their usage in music or percus-
sive sound analysis. In particular, the percussive content of music generally
exhibits repeating patterns. Even though they contain the same sequential
data within the musical piece, the patterns tend to vary between pieces. As
a result, temporal prediction operating on immediately preceding events may
not be the most efficient way to model repeating rhythmical patterns.
Based on the above observation, Paulus and Klapuri proposed the use of
periodic A^-grams [506] where, instead of using the directly preceding A^ 1
events, the idea is to take the earlier events separated by an interval L. That is,
when predicting the event at temporal location k. instead of using the events
at locations k N + l^k N-[-2,...,k 1. use the events at the locations
k-{N - 1)L, k-{N - 2 ) L , . . . , A: - L. The A^-gram model of (5.23) is then
reformulated as

p{Wk\wi:k~l) = P(^A;k/c-(;V-l)L^/c-(7V-2)L U)k-L)- (5.26)

It is easy to see that by assigning L 1, the model reduces to the standard

A^-gram of (5.23). The repeating patterns tend to occur at integer multiples
of the musical-measure length of the the piece. Due to this, it was suggested
that the interval L should be set to correspond to the measure length A of the
piece under analysis.
The idea of the normal and the periodic A^-grams are illustrated in Fig. 5.6.
The horizontal arrow represents the use of a normal trigram (A^-gram of length
3) to predict the event at the location of the question mark, whereas the
158 Deny FitzGerald and Jouni Paulus
tatum 1 2 3 4 5 6 7 8
isure 1 BH H H H BH H H H

isure 2 BH H H H @ HC SH HT

isure 3 BHC H SH H ^ H SH H

isure 4 BH H @~ -(uy"*"?\
Fig. 5.6. The idea of the normal and periodic A^'-grams illustrated. Time is quantized
to a tatum grid, each box representing a segment between two grid points. Time
flows from left to right continuing on the next row, each row being one musical
measure. The letters represent the drum instruments played at the corresponding
time instants (B is kick drum, S is snare drum, H is hi-hat, T is tom-tom, and C
is cymbal). The horizontal arrow represents a normal trigram prediction, and the
vertical arrow represents a periodic trigram prediction. The measure length here is
eight tatum periods and L = 8 is set accordingly.

vertical arrow illustrates the use of a periodic A^'-gram. It should be noted

that both the TV-gram types can be constructed either for mixture events or
individual instruments. The latter type of models can be visualized by looking
at only one hit type (letter) at a time. For example, when estimating the
probability of a hi-hat event occurring at the position of the question mark,
the model observes only that the two preceding mixtures contain a hi-hat hit.
Use of different A^-gram models was evaluated in [506]. The low-level analy-
sis was done by creating a tatum grid over the signal, based on manually
annotated information. Then a set of features was extracted at each seg-
ment between two grid points, and a combination of GMMs and an SVM was
used to recognize its contents. A separate GMM was trained for each possible
mixture-event, and an SVM was used as a silence detector. At each segment,
the low-level recognition produced likelihoods for diflPerent mixture events. The
actual transcription was done simply by choosing the mixture having the high-
est likelihood at each grid point. The acoustic recognition result was relatively
poor, with an instrument event error rate of 76%. All the high-level process-
ing methods were added on top of this. On adding the mixture-event prior
probabilities, the error rate dropped to 50%. With mixture-event bigrams
and trigrams, the periodic version performed slightly better than the stan-
dard ones. Also, the longer A^-grams for individual instruments with N = 5
and A^ = 10 performed better than the shorter mixture-event A''-grams. The
best result, obtained with traditional non-periodic A/^-grams for individual in-
struments using N = 10, still had a 46% error rate. This demonstrates that
musicological models cannot greatly improve the overall result if the low-level
recognition results are poor to start with. Instead, they require suflSciently
accurate information to be available from low-level recognition to be able to
make proper assumptions about the musical structure and improve the overall
result.
 5 Unpitched Percussion Transcription 159

5.4.4 Modelling of Compositional Rules

Musicological models have been most commonly used in connection with su-
pervised classification methods, such as GMMs and HMMs, but they can also
be employed with unsupervised methods, such as event clustering. This was
done by Paulus and Klapuri, who used a musicological model to label per-
cussive sound events with instrument names of the regular rock/pop drum
kit, even though the input signal was produced using arbitrary sounds [504].
In other words, the input signal was a percussive sequence created with non-
drum sounds, like speech sounds or tapping different objects, and the output
transcription was intended to map these to a normal drum kit. Because the
labelling or mapping could not be determined from the acoustic properties
of the sounds, it was done with a simple musicological model of drum loops.
The model consisted of the probabilities of different percussive instruments to
be present at different temporal locations within a musical measure. This re-
quires that the boundaries of musical measures are located in the input signal,
which can be done with the methods discussed in Chapter 4.
The system presented by Paulus and Klapuri [504] operated with the fol-
lowing steps. First, temporal segmentation was carried out by onset detection.
Acoustic features were extracted at each segment and, based on these, the seg-
ments were clustered into K categories. The onset times were then quantized
to an estimated tatum grid. At this point, each tatum grid point contained
information as to which cluster the sound event at that point belonged to.
Finally, suitable labels for the clusters were determined so that the generated
sequence would make sense when performed with a drum kit.
The above labelling problem was formulated as finding a mapping M from
the K clusters to the available labels M: {0,1, 2 , . . . , K } -^ { 0 , 5 , 5 , i 7 } ,
where the labels denote silence, kick drum, snare drum and hi-hats, respec-
tively.^ As the total number of different mappings was relatively small, all
of them could be tested and the best one chosen. The likelihood of a certain
mapping M was evaluated with

p{M) = Wp{qi\n,,\),^ (5.27)

i

where p{M) is the likelihood of the mapping, and p{q\n^ A) is the probability
of the label q to be present at the temporal location n G { 0 , . . . , A 1} when
the length of the musical measure is A. The total likelihood is calculated over
the whole signal containing all the events z.
The system was evaluated with acoustic signals synthesized from a com-
mercial MIDI database comprising a wide variety of different percussive tracks
[305]. The synthesis was done by using sampled speech sounds and the sounds
of tapping different objects in an ofRce environment. There were fifteen sam-
ples for each sound type, and each synthesized hit was randomly selected
^In the general case, K does not need to be equal to the number of available
labels.
160 Derry FitzGerald and Jouni Paulus

from this set to produce realistic acoustic variation to the synthesis result.
The overall error rate of the system was 34%. Error analysis revealed that
there were large differences in performance between different genres, and the
genres with simpler rhythmic patterns were labelled more accurately.
It has been established that musicological modelling is useful in the con-
text of percussive sound transcription. However, the low-level analysis has to
be done sufficiently accurately before the musicological modelling can really
improve the results obtained. Further, the methods required to combine low-
level acoustic recognition with the high-level modelling still need development.
As percussive patterns tend to be different in different time signatures, styles,
and genres, specific models for each of these could be developed.

5.5 Conclusions
An overview of the current state of the art in unpitched percussion transcrip-
tion has been presented. This encompassed both event-based and separation-
based systems, as well as efforts to include high-level language modelling to
improve system performance. As can be seen, there has been considerable
effort expended on tackling the problem of percussion transcription in the
past few years, and a summary of the important systems to date is presented
in Table 5.1.
At present, the best performance has been obtained on systems that focus
on a reduced number of drums: snare, kick drum, and hi-hats in the drums-
only case, and snare and kick drum in the presence of pitched instruments.
This is unsurprising in that the complexity of the problem is greatly reduced
by limiting the number of target instruments. Nonetheless, these systems do
deal with the most commonly occurring drums, and so represent a good start-
ing point for further improvements.
As noted above, many of the systems do not take into account the pre-
dictability of percussion patterns within a given piece of music. However, it
has been established that the use of musicological modelling does consider-
ably improve the performance of a system using only low-level processing. In
particular, it should be feasible to integrate musicological modelling to many
of the separation-based models.
There has also been a trend towards adaptive systems that take into
account the characteristics of the signals being analysed when attempting
transcription, both in event-based and separation-based systems. This is an
attempt to overcome the large variances in the sounds obtained from a given
drum type such as a snare drum. For example, drums in a disco-style genre
have a totally different sound to those in a heavy metal-style piece. These
adaptive systems are to be encouraged, as a system that can be tailored to
suit individual signals is more likely to produce a successful transcription than
a system which makes use of general models.
 5 Unpitched Percussion Transcription 161

Table 5.1. Summary of percussion transcription systems. The column Classes con-
tains the number of percussion classes covered by the system. Method describes
the overall approach, E for event-based systems, S for separation-based systems,
M for systems including musicological modelling, and A for systems using adaptive
modelling. An X in Drums only indicates that a system operates on signals con-
taining drums only. Mono/Poly shows whether the system can detect two or more
simultaneous sounds

Authors Classes Drums Mono/ Method Main algorithms

only Poly
Dittmar & Uhle [147] 5 P S+A Non-negative ICA
FitzGerald et al. [187] 3 X P S Sub-band ISA
FitzGerald et al. [188] 3 X P S+A PSA
FitzGerald et al. [189] 3 P S PSA
FitzGerald et al. [190] 7 X P S-hA Adaption & PSA
Gillet & Richard [209] 8 X P E+M HMM, SVM
Goto & Muraoka [235] 9 X P E Template matching
Goto [221] 2 P E Frequency histograms
Gouyon et al. [250] 2 X m E Feature extraction
Herrera et al. [287] 9 X m E Various
Herrera et al. [285] 33 X m E Various
Paulus & Klapuri [506] 7 X P E+M GMMs & A^-grams
Paulus & Virtanen [505] 3 X P S NMF
Sand void et al. [561] 3 P E+A Localized models
Sillanpaa et al. [590] 7 P E+M Template matching
Van Steelant et al. [608] 2 P E SVMs
Virtanen [650] 2 P S Sparse coding
Yoshii et al. [682] 2 P E+A Template matching
Zils et al. [693] 2 m E+A Template matching

A problem with the research on percussion transcription to date has been

the lack of comparability of results. This is due to varying problem formula-
tions, and to the use of disparate test sets and evaluation measures. T h e two
different comparative evaluations mentioned above (see pp. 148 and 149) are
among the few t h a t have compared different systems on the same test data.
Recently, however, a number of different transcription systems were evaluated
in the framework of the Audio D r u m Detection contest^ organized in connec-
tion with the 6th International Conference on Music Information Retrieval.
T h e contest is a clear step towards an extensive and regular comparative
evaluation of different methods.
Given the noticeable progress in percussion transcription in the past few
years, it is hoped t h a t the future will see a further improvement in the per-
formance of the systems, and, as the performance of the systems improves,
in the range of percussion instruments covered. T h e problem is still far from

^The results are available at www.music-ir.org/mirexwiki.

162 Derry FitzGerald and Jouni Paulus

solved, and it is hoped that this chapter reflects only the beginning of the
study of unpitched percussion transcription.

5.6 Acknowledgements

Derry FitzGerald was supported in this work by the Irish Research Council
for Science, Engineering and Technology.
6
Automatic Classification of Pitched Musical
Instrument Sounds

Perfecto Herrera-Boyer/ Anssi Klapuri,^ and Manuel Davy^

^ Institut Universitari de I'Audiovisual, Universitat Pompeu Fabra,

Pg. CircumvaMacio 8, 08003 Barcelona, Spain
pherreraQiua.upf.es
^ Institute of Signal Processing, Tampere University of Technology,
Korkeakoulunkatu 1, 33720 Tampere, Finland
[email protected]
^ LAGIS/CNRS, BP 48, Cite Scientifique, 59651 Villeneuve dAscq Cedex, France
Manuel.DavyQec-lille.fr

6.1 Introduction
This chapter discusses the problem of automatically identifying the musical
instrument played in a given sound excerpt. Most of the research until now
has been carried out using isolated sounds, but there is also an increasing
amount of work dealing with instrument-labelling in more complex music sig-
nals, such as monotimbral phrases, duets, or even richer polyphonies. We first
describe basic concepts related to acoustics, musical instruments, and percep-
tion, insofar as they are relevant for dealing, with the present problem. Then,
we present a practical approach to this problem, with a special emphasis on
methodological issues. Acoustic features, or, descriptors, as will be argued, are
a keystone for the problem and therefore we devote a long section to some of
the most useful ones, and we discuss strategies for selecting the best features
when large sets of them are available. Several techniques for automatic classi-
fication, complementing those explained in Chapter 2, are described. Once the
reader has been introduced to all the necessary tools, a review of the most rele-
vant instrument classification systems is presented, including approaches that
deal with continuous musical recordings. In the closing section, we summarize
the main conclusions and topics for future research.

6.1.1 Classification in Humans and Machines

Classification, in practical terms, refers to the process of assigning a class label

to a given observation. This observation is typically described as a numerical
vector that represents some features of the observation. In the case of mu-
sic analysis systems, the features are computed from the raw audio signal by
164 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

means of signal processing techniques. Class labels are by definition categori-

cal data types representing some kind of generalization about the observable
world. They are usually organized by means of structures, such as taxonomies
and ontologies, that exhibit the following ideal features: (1) consistency in
the principles of classification, (2) mutual exclusiveness of categories, and (3)
completeness, i.e., total coverage of the universe that they attempt to describe
[46].
The interest in automatic classification of musical instrument sounds is
manifold:
From the acoustics point of view, to understand what makes the sound of
a given instrument 'identifiable' among other instruments;
From the perceptual perspective, to understand what makes the sounds of
two different instruments 'similar';
From the sound hbrarian perspective, to automatically provide labels for
retrieving desired items from sound sample databases or synthesizer li-
braries;
From the musicological point of view, to locate solo sections or the ap-
pearance of a certain instrument in a musical recording.
Classification learning can proceed in a supervised way, where the system
is provided with training data and the corresponding labels that should be
associated with them, or in an unsupervised way, where there is no label to be
associated with a given example but the system groups objects according to
some similarity or homogeneity criteria. In the latter case, we are effectively
doing clustering^ a task that is slightly different from classification [674].
The supervised classification procedure can in general be described as fol-
lows (see Fig. 6.1):
1. Acoustic features are selected to describe sound samples.
2. Values of these features are computed for a labelled training database.
3. A learning algorithm that uses the selected features to learn to discrimi-
nate between instrument classes is applied and fine tuned, exploiting the
training database.
4. The generalization capabilities of the learning procedure are evaluated by
classifying previously unseen sound samples (cross-validation).
Research on the automatic classification of musical instruments sounds
has focused, for a long time, on classifying isolated notes from different in-
struments. This is an approach that has a very important trade-off: we gain
simplicity and tractability, as there is no need to first separate the sounds
from a mixture, but we lose contextual and time-dependent cues that can
be exploited as relevant features when classifying musical sounds in complex
mixtures.
Human discrimination of sound sources is based on spectral and tempo-
ral cues that are extracted by the early processing in the cochlea and up
to the primary auditory cortex, whereas class decisions seem to be made in
 6 Automatic Classification of Pitched Musical Instrument Sounds 165

TmiraiYgSel

Test St

'K^\l^^
Fig. 6 . 1 . Diagram of the different operations involved in setting up an automatic
classification system for musical instrument sounds. The training set is described by
extracting an initial set of features that is refined by means of feature transformation
and selection algorithms. The resulting set of selected features is used to train and
validate (i.e., fine-tune) the classifier. Validation can be done using a different set
of sounds (not shown here) or different partitions of the training set. When fine-
tuning is finished, the classifier is tested with a test set of sounds in order to assess
its expected performance. At the right side of the diagram, using dotted elements,
the automatic classification of an unlabelled (i.e., previously unseen) sound file is
illustrated.

higher auditory centres [260]. In the case of pitched musical instruments, the
relative strengths of the overtone partials (see Fig. 6.2) determine, to a cer-
tain extent, timbre sensations and identification. It seems t h a t , for sustained
sounds, the steady segment provides much more information t h a n the attack,
though the latter cannot be completely neglected [270]. Timbre discrimina-
tion experiments, where sounds are altered in subtle or dramatic ways and
the listeners indicate whether two different versions sound the same, have
provided cues concerning the relevant features for sound classification. Grey
and Moorer [254] found t h a t microvariations in amplitude and frequency are
usually of little importance, and t h a t the frequency and amplitude envelopes
can be smoothed and approximated with line segments without being noticed
by the listeners. Changes in temporal parameters (i.e., attack time, modula-
tions) may have a dramatic impact on the discrimination of timbres [83], [560],
166 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

1000 2000 3000 4000 5000 6000

2000 3000 4000

Frequency (Hz)

Fig. 6.2. Example spectra from three different instruments. Prom top to bottom:
a clarinet (playing the note C4), a violin (C6) and a guitar (G5). In the clarinet
sound, note the predominance of the first even partials. In the guitar sound, note
the existence of plucking noise as energy unevenly distributed below 700 Hz.

though probably not in deciding the instrument name. The above facts were
also supported by McAdams et al. [446], who, additionally, found that the
spectral envelope shape and the spectral flux (time-variation of the spectrum;
see Section 6.3.2) were the most salient physical parameters affecting tim-
bre discrimination. Also, it has been noticed that the human sensitivity for
different features depends on the sound source in question.
Very few studies have investigated the human ability to discriminate be-
tween the sounds of different musical instruments. However, some trends can
be identified based on the reviews and experiments by Martin [442] and
Srinivasan et al. [606]. First, humans, even those with musical training, rarely
show performance rates better than 90%. The number of categories in the cited
experiments varied from 9 to 39, and in the most difficult cases the recognition
rate dropped to 40%. Second, confusion between certain instruments are quite
usual, for example, between the French horn and the trombone. Third, the
discrimination performance can be improved by musical instruction and by
exposure to the acoustic material, especially to pairs of sounds from different
instruments. Fourth, instrument families are easier to identify than individ-
ual instruments. Finally, contextual information (i.e., listening to instruments
playing phrases, instead of isolated notes) substantially improves the identifi-
cation performance [340], [57].
 6 Automatic Classification of Pitched Musical Instrument Sounds 167

Table 6.1. A simplified taxonomy of pitched musical instruments that is partially

based on Hornbostel and Sachs [296].
Family Subfamily Playing mode Instrument
Chordophones Bowed Contrabass, Violoncello, Viola, VioHn
Plucked Guitar, Harp, Ukulele, Harpsichord
Struck Piano
Aerophones Woodwind Air reed Flute
Double reed Oboe, English Horn, Bassoon
Single reed Saxophone, Clarinet
Brass With valves Tuba, French Horn, Trumpet
W / o valves Trombone, Bugle
Idiophones Struck Xylophone, Marimba, Celesta, Bell
Plucked KaUmba, Mbira
Membranophones Struck Timpani

6.1.2 Taxonomies of Musical Instruments

Taxonomies are a way to organize or classify musical instruments from the

point of view of organology, which is the discipline dealing with the cultural,
historical, technological, and practical issues of musical instruments [329].
When building taxonomies, we generate a tree where each category-branch
can be broken down into more detailed categories. As we will see below,
taxonomies are an important tool not only for conceptualizing similarities
and differences between instruments, but also for helping our algorithms to
discriminate between classes of sounds based on their acoustic properties.
Taxonomies can be also termed classification schemes in certain technical
domains, such as MPEG-7 [432].
One of the most extended and exhaustive taxonomies of musical instru-
ments is that of Hornbostel and Sachs [296], which considers instruments
according to how their sound is produced (see Table 6.1). Using this criterion,
instruments are divided into chordophones (instruments that produce sound
by acting on strings), aerophones (instruments that produce sound by acting
on a wind column), idiophones (instruments that produce sound by acting on
their own bodies), and membranophones (instruments that produce sound by
acting on an elastic membrane). A recently added category is that of elec-
trophones, which includes instruments that produce their sound by electrical
means, such as electronic organs and synthesizers.
An additional taxonomic distinction to complement that of Hornbostel and
Sachs comes from considering pitch: there are instruments that cause a clear
pitch sensation (chordophones and aerophones), whereas the sounds of some
other instruments do not have a definite pitch (most of the idiophones and
membranophones). We denote the former ones as 'pitched', 'tuned', or 'with
determinate pitch', whereas the latter ones are considered 'unpitched', 'un-
tuned', or 'with indeterminate pitch'. Only pitched instruments are shown in
168 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

Table 6.1. Other criteria for elaborating subclasses can be the playing method,
the shape of the instrument, the relationships of the exciting element to the
resonating element, or the method used to put the exciting element into mo-
tion. For a detailed account on the acoustics of musical instruments, the reader
is referred to Fletcher and Rossing [193] and Rossing [551].

6.1.3 Timbre Spaces

The perceived similarity or dissimilarity of the sounds of different musical

instruments can be characterized with an abstract representation called 'tim-
bre space'. Timbre spaces reflect human perception and are not necessarily
optimal from the viewpoint of partitioning the space into separable classes.
However, they can reveal the acoustic properties that enable computing per-
ceptual similarities between instrument classes.
In order to derive timbre spaces, we first have to ask human listeners
to generate similarity judgements by listening to pairs or triads of sounds.
Similarity is recorded using a continuous or discrete scale that goes from
'very similar' to 'very dissimilar'. Similarity scores are then processed using
a data dimensionality reduction technique known as multidimensional scal-
ing (MDS). MDS is a method that represents measurements of dissimilarity
between pairs of objects as distances between points in a space with a small
number of dimensions (usually two or three), where each object (here, sound)
corresponds to a point within the space. The MDS algorithm is explained in
detail in Section 6.3.5.
When using timbre spaces to represent similarities between sounds, the
obtained continuous space is delimited by axes that correspond to perceptual
or acoustic factors having an important role in defining the timbre sensation.
By looking at the correlations between the axes and the acoustic features
that are computed from the sounds evaluated by the listeners, meaningful
interpretations of the space can be derived.
Grey [253] was one of the pioneers in the elaboration of timbre spaces. He
used synthetic sounds emulating twelve orchestral instruments from the string
and the wind families. After requesting twenty listeners to judge the similar-
ity between pairs of sounds, he applied MDS to the similarity judgements to
derive a timbre space with three dimensions. The qualitative description of
the axes he obtained when looking at their acoustic correlates were: (1) the
spectral energy distribution, (2) the amount of synchronicity in the begin-
ning transients, and (3) the temporal variations in the spectral envelope of a
sound.
Further research by Wessel [665], Krumhansl [378], Krimphoff et al. [374],
McAdams et al. [448], and Lakatos [385] have supported the predominance of
a 'brightness' dimension (which is related to the first moment of the spectral
energy distribution) as the main perceptual attribute for organizing sounds
in a timbre space. Another important attribute seems to be related to the
attack time, and a third dimension, receiving less clear support, could be the
 6 Automatic Classification of Pitched Musical Instrument Sounds 169

'spectral flux', which is related to the temporal variation of the spectrum.

These and other related features were included for timbre description in the
MPEG-7 standard (see Section 6.3) [513].
Although most of the research on timbre spaces has been done in order
to characterize human perception tendencies, it is also possible and useful to
build them from purely acoustic measures. In this case, they reflect similarities
inherent to the feature vectors that are used to characterize the sounds [105],
[285], [510], [419].

6.2 Methodology

In this section, we discuss the methodological issues involved in building and

evaluating classifiers which learn classes from labelled training data.

6.2.1 Databases

Databases are one of the crucial elements needed for developing a successful
classification system, as they have to include enough 'representative examples'
in order to grant the generalizability of the models built upon them [112].
Proper data modelling requires the careful preparation of up to three different
and independent data sets (see Fig. 6.1). The first one, usually termed a
training set^ is used to build the models, whereas the second one, usually
termed a testing set^ is only used to test the model (or the system using it)
and to get an estimate of its efficacy when it will be running in a real-world
system. A third set, usually termed a validation set, is sometimes used during
the design, improvement, and tweaking of a given model. In that case, the
model, as it evolves and improves, is tested using the validation set, and only
when the model preparation phase is finished (i.e., when the performance
improvement on the training data is no longer matched by the performance
improvement on the validation set) is it evaluated against the testing set,
which is kept untouched until then. Of course, the three sets should be sampled
from the same population of sounds.
A testing set that could be shared among research teams would help them
to compare their respective improvements. Unfortunately, most of commer-
cial audio files, MIDI files, and digitalized score files cannot be shared. In the
automatic classification of musical instruments, the commercial McGill Uni-
versity Master Samples collection (MUMS) [487] has been frequently used,
though it has not achieved the status of 'reference test set'. More recently,
the University of Iowa sample collection^ and, especially, the RWC database
[230] are attracting the attention of researchers. The latter contains a wide
variety of music files to be used in several music processing problems, and

^http://theremin.music.uiowa.edu/index.html
170 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

it is available to researchers under reasonable restrictions. Also, the Interna-

tional Music Information Retrieval Systems Evaluation Laboratory [154] is
intended to provide remote access to a protected commercial music library
and to a computational infrastructure that hosts algorithms for music con-
tent processing. In this way evaluation and comparison of different algorithms
is possible, as the data sets and the computer systems are kept invariable. A
simpler option, though, is now possible, thanks to the Creative Commons^
initiative, which is a scheme for licensing and managing intellectual property
rights. Several music providers are running their businesses partly or entirely
on Creative Commons and therefore the music they distribute can be 'shared'
by the scientific community.

6.2.2 Validation and Generalization

When we develop a classification system that learns from labelled examples, it

is necessary to assess its generalization capabilities, i.e., its performance when
classifying previously unseen instances. This process is called validation, and
there are different procedures that can be followed: the leave-one-out method,
the TV-fold cross-validation, and the independent or holdout test set. The first
one, recommended when we have a small number of instances available, tests a
system with the same instances that are used to learn the classification model:
If there are K instances in the data set, then K different models are built,
each time leaving one instance out from the training and keeping it apart to
test the model [379]. The results are then averaged over the K tests. In the N-
fold cross-validation, we randomly divide the data set into N subsets or folds
(usually 10) and then use N 1 folds for training and one for testing. This is
repeated A^ times and the final result is again averaged over the N runs (see
also the alternatives and recommendations issued by Dietterich [145]). When
a very large set of labelled instances is available, the recommended validation
procedure consists of using a holdout set that has been kept unused up until
then. Livshin and Rodet [414] have also presented interesting variations that
can be used when we have several large sound collections available.
Another methodological element is the assessment of the statistical signifi-
cance of the differences that have been found when comparing different sets
of features or different classification algorithms. Empirical differences in per-
formance can be obtained by chance or, alternatively, can be the consequence
of true substantial differences in the goodness of the compared methods. The
Student-Fisher t-test is frequently used to compare the error averages of dif-
ferent classification strategies [145]. It can be computed, in the case of A^-fold
cross-validation, as
p^/N
t= I (6-1)

^http://creativecommons.org
 6 Automatic Classification of Pitched Musical Instrument Sounds 171

where N is the number of folds, p^*^ is the average at fold i, and p =

'k 12i=i P^^"^' Under the null hypothesis (i.e., the hypothesis that the aver-
age performances of two methods are not different) this statistic has a t-
distribution with n 1 degrees of freedom. The t-distribution is a symmetric
bell-shaped distribution with the tails spread wider than those of the normal
distribution. Its exact shape depends on the degrees of freedom: the greater
the sample size, the more degrees of freedom and the narrower the bell shape
(a large sample size makes the ^-distribution resemble the normal distribu-
tion). The significance of the value t in (6.1) can be evaluated by looking it
up in a table of the t-distribution and, based on this, the null hypothesis can
be either rejected or accepted.
In spite of the pervasiveness of ^-tests, it is sometimes recommended that
one use the McNemar's test instead, as it provides a greater chance of ob-
taining a statistically significant result when the null hypothesis is false, and
thus gives more power to our conclusions. The value for McNemar's test is
computed as
^ 2 ^ (|moi - m i o | - l ) ^
moi -h mio
where moi is the number of instances misclassified by algorithm A but not by
algorithm B, and mio is the number of instances misclassified by algorithm B
but not by algorithm A [145]. Again, the significance of the obtained value has
to be evaluated by looking at its distribution table, and this tells us whether
to reject the null hypothesis or not.
Other methodological issues to be considered are the scalability of the al-
gorithms, or their efficiency (i.e., what happens to the computation time when
we double the number of features or the number of instances to be processed?).
For methodological issues in computer experiments, the interested reader is
referred to Santner et al. [562].

6.3 Features and Their Selection

Classification would be not possible without acoustic features that capture
differences or similarities between two or more objects. In the case of auto-
matic classification of instrument sounds, the features are extracted from the
raw audio signal. As we will see, choosing good features is more crucial than
the choice of the classification algorithm, and the classification itself becomes
easier if the features chosen are informative enough [216].
The term 'feature' denotes a quantity or a quality describing an object.
Features are also known as attributes or descriptors. Since the overall goal
of classification is to distinguish between examples that belong to different
classes, one of the goals of feature extraction is to reduce the variability of
feature values for those examples that are associated with the same class, while
increasing the variability between examples that come from different classes.
Another informal requirement is to achieve a good capability of predicting the
class after observing a given value of the feature (i.e., a high 'diagnosticity').
172 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

Finally, when using multiple features to predict a class label, it is desirable

that the features be as little redundant as possible. Knowledge about the
acoustic and perceptual qualities of the specific classes to be discriminated
should guide, if possible, the choice of the features to use. Even lacking such
knowledge, however, powerful feature selection and projection algorithms can
be used to improve the discrimination capability.
Acoustic features are usually computed directly from the time-domain sig-
nal or, alternatively, from a representation obtained by transformations such
as the discrete Fourier transform (DFT) or the wavelet transform, which have
been discussed in Chapters 2 and 3, respectively. In order to grasp the micro-
temporal evolution of the feature values, they are typically calculated every
few milliseconds, in short analysis windows. Features representing the prop-
erties of instrument sounds are typically extracted within frames of length
20-50 ms, with a 50-75% overlap between successive frames. Macro-temporal
features, such as attack time or vibrato rate, can be computed by using
a longer segment of audio or by summarizing micro-temporal features over
longer segments (for example with averages and variances).
For the purpose of presentation, in the following we have organized the
described features under the subtitles of energy features, spectral features,
temporal features, and harmonicity features (see [286], [511], [406], [694] for
other feature taxonomies). As more than one hundred different features have
been tested in the context of our problem, we only present the ones that con-
stitute a basic kit to start with (see [363], [511] for more extensive accounts).
An interesting set of features that has recently received some attention is the
one included in MPEG-7, an ISO standard for multimedia content description
(see Table 6.2) [432]. The interested reader is referred to [342] for in-depth
explanations of the MPEG-7 audio descriptors and applications. These de-
scriptors have been tested for the classification of musical instrument sounds
in [71], [615], [367].

6.3.1 Energy Features

Signal energy or power can be measured at different time scales and used as
an acoustic descriptor. Although these have not shown a high discriminative
power when compared to other features used for musical instrument classifi-
cation, they provide basic information that can be exploited to derive more
complex descriptors, or to filter out potential outhers of a sound collection.
The root mean square (RMS) level of a signal is often used to represent
the perceptual concept of loudness. The RMS level of a discrete time signal
x{n) is calculated as

1 ^'^

\ 2= 0

where n is a discrete time index and A^ is the size of the analysis frame.
 6 Automatic Classification of Pitched Musical Instrument Sounds 173

Table 6.2. MPEG-7 descriptors, organized according to category. Each of these

can be used to describe an audio segment with a summary value or with a series
of sampled values. Timbral spectral descriptors are computed after extracting the
relevant harmonic peaks from the spectrum of the signal. Spectral basis descriptors
are a spectral representation of reduced dimensionality.
Category Descriptors
signal parameters fundamental frequency, harmonicity
basic instantaneous waveform, power values
basic spectral log-frequency power spectrum envelopes, spectral centroid,
spectral spread, spectral flatness
timbral spectral harmonic spectral centroid, harmonic spectral deviation,
harmonic spectral spread, harmonic spectral variation
timbral temporal log attack time, temporal centroid
spectral basis spectrum basis, spectrum projection

6.3.2 S p e c t r a l F e a t u r e s

Different approaches can be considered in order to measure the shape of the

spectral envelope or certain characteristics of the spectral fine structure. Spec-
tral features can be computed using the D F T spectrum, or by applying a
mel-scale filterbank and using the RMS levels at the mel bands. In order t o
do the latter, we first construct a bank of about 40 filters t h a t are spaced uni-
formly on the mel frequency scale and have triangular magnitude responses
(see Fig. 2.4 on p. 27). RMS levels within each mel band are then calculated
by multiplying the magnitude spectrum of a given analysis frame with each
triangular response, by squaring and summing the resulting D F T bin magni-
tudes, and by taking the square root of the result. T h e mel frequency scale
approximates the way humans perceive pitch, having less resolution at high
frequencies and a finer resolution at low frequencies. It can be computed as

/mel = 2595 logio (6.4)

700

Another psychoacoustically motivated frequency scale is the Bark scale, where

usually 24 sub-bands are used.
A selection of the most useful spectral features is presented in the following.
T h e spectral centroid, spectral skewness, spectral kurtosis, and spectral spread
have been described in Chapter 5, p. 136, and are therefore not included here.
In t h e presented equations, X{k) may refer to a D F T spectrum or to a vector
of RMS levels in Mel bandsdepending on the pre-processing t h a t has been
adoptedand accordingly, the symbol k may index a frequency bin or a mel
sub-band, and B denotes the total number of frequency bins or sub-bands
used.
Spectral flatness indicates how flat (i.e. 'white-noisy') the spectrum of a
sound is. A low value indicates a noisy sound, whereas a high value is indicative
174 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

of tonal sounds. It is computed as the ratio of the geometric mean to the

arithmetic mean of the spectrum:

(ntil^t(fc)l)"
SFM(^) = 101ogio (6.5)

where t is the index of the analysis frame.

The spectral flux, also known as the delta spectrum magnitude, is a mea-
sure of local spectral change. It is defined as the squared l'^ norm of the
frame-to-frame spectral difference:

SFX{t) = f2{\Mk)\-\XtMk)\)\ (6.6)

k=l

where Xt(A:) and Xt-i{k) are energy-normalized Fourier spectra in the current
frame and in the previous frame, respectively.
Spectral irregularity measures the 'jaggedness' of the spectrum [374]. It is
computed as
(6.7)

Spectral roll-off is defined as the frequency index R below which a certain

fraction 7 of the spectral energy resides:

|X(A:)|2>7f^|X(fc)|2 (6.8)
k=l k=l

Typically, either 7 = 0.95 or 7 = 0.85 is used.

The Zero crossing rate (ZCR) measures the number of times that the time-
domain signal changes its sign. Even though it is computed in the time domain,
it describes the amount of high-frequency energy in the signal (i.e., 'bright-
ness') and correlates strongly with the spectral centroid mentioned above.
ZCR has also proven to be quite discriminative for classes of percussion in-
struments [250]. The ZCR of a time-domain signal x{n) is computed as

1 ^
ZCR(n) ^w^J2 Isign[x(n + i)] - sign[a:(n + i - 1)]|, (6.9)
i=l

where
-hi ifx>0
sign(a:) = < 0 if X = 0 . (6.10)
-1 ifx<0
 6 Automatic Classification of Pitched Musical Instrument Sounds 175

Mel-frequency cepstral coefficients (MFCC) were introduced in Chapter 2,

p. 26. To summarize the computational procedure: (1) a pre-emphasis filter is
applied to flatten the spectrum; (2) audio is windowed and transformed using
the DFT; (3) a mel-scale filterbank is applied in the frequency domain, as
explained above; (4) the power within each sub-band is computed by squaring
and summing frequency bin magnitudes within bands; (5) the dynamic range
of the spectrum is compressed by taking a logarithm of the bandwise power
values; and (6) cepstral coefficients are computed by applying, to the log
filterbank powers, the discrete cosine transform (DCT) which decorrelates
the coefficients. The dimensionality of the representation can be reduced by
retaining only approximately 15 lowest-order DCT coefficients, which usually
carry the relevant timbral information.
MFCCs have proven to be useful not only in speech processing tasks [536]
but also in instrument recognition as shown, for example in [675], [56], [437],
[176], [178]. Sometimes the delta ('velocity') and delta-delta ('acceleration')
coefficients are also used, in order to capture information about their tem-
poral evolution. The delta-MFCCs are usually computed as a least-squares
approximation to the local slope, or

^Cep.(,)=^"-ir'^'""/'', (6^.1)

where t is the frame index, Cep^(f) is the ith coefficient in frame t, and usually
M is 1 or 2. The delta-delta, in turn, can be computed by substituting Cep^{t)
by Z\Cep^(t) in the above equation.

6.3.3 Harmonic Features

Pitched instruments usually have harmonic or nearly harmonic spectra. There-

fore, it is interesting, as reported in several papers [369], [199], [688], [514],
[670], [495], to use descriptors that take this property into account. Different
options exist to extract the harmonic partials from a signal, such as computing
the autocorrelation of the signal and generating an autoregressive (AR) model,
selecting prominent peaks at integer multiples of the fundamental frequency
from a long-term average spectrum, applying pattern-matching techniques to
the peaks of the DFT spectrum, or estimating a sinusoid plus noise model
[574] (see also Chapter 3). In this section, we denote the amplitude of the ^th
partial by a{q), and the number of extracted partials by Q.
Inharmonicity describes the average deviation of spectral components from
perfectly harmonic frequency positions. We can compute it as

IH = -ET X _Q ^7^ , (6.12)

ZU^'ii)
176 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

where fq is the frequency of partial g, and FQ is the estimated fundamental

frequency. An inharmonicity value of zero means that the signal is perfectly
harmonic, whereas larger values indicate inharmonicity.
The odd-to-even ratio describes the relationship between odd and even
partials. Some instruments, such as clarinets, have odd partials that are more
prominent than the even ones. Other instruments, such as trumpets, have
more balanced spectra. The odd-to-even ratio is given by

OER=^^i^-^^. (6.13)
E g even ( ^ )

Tristimulus consists of three features that describe the relative weights of

different harmonics [527]. These are computed as follows:

Tl = -P^^, (6.14)

^ ^ ^ a ^ ( 2 ) + a ^ ( 3 ) + a-(4)

By plotting the temporal evolution of T2 (x-axis) against T3 (?/-axis),

we can visualize the relative strengths of the fundamental and the high- and
mid-frequency partials, and the evolution of these through time (see Fig. 6.3).

6.3.4 Temporal Features

The time dimension is usually less represented in the feature sets proposed for
the automatic classification of musical sounds. The evolution of a given feature
over time can be partially characterized by computing its variance, or the first-
and second-order differences. Apart from these, specialized descriptors, such
as the attack time, the temporal centroid, or the rate and depth of frequency
modulation have proven to be useful for discriminating between instrument
sounds [442], [199], [514], [615], [367].
The term amplitude envelope is generally used to refer to a temporally
smoothed version of the signal level as a function of time. In practice, it can
be calculated by lowpass filtering (with a 30-Hz cut-off frequency) the vector
of RMS levels E{n) of a signal. In the case of analysing isolated notes, once
the envelope is computed, it is possible to segment it into attack, sustain, and
release sections, as shown in Fig. 6.4 (though percussion and plucked string
sounds do not have the sustain part). Specific descriptors for each of these
segments can also be computed.
 6 Automatic Classification of Pitched Musical Instrument Sounds 177

Strong
mid-frequency
partiais

Strong
high-frequency
partiais

Fig. 6.3. Geometric interpretation of the tristimulus. The left figure shows the re-
gions where, depending on the energy balance, the values of of T2 and T3 will be
found. The figure on the right illustrates the temporal evolution (in milliseconds)
of a clarinet note: it starts with a strong fundamental, then high frequency par-
tiais progressively dominate the sound, and finally, after 60 milliseconds, the high
frequencies start to decay until the end of the sound.

Attack

Fig. 6.4. SimpHfied amplitude envelopes of a guitar tone (above) and a violin tone
(below). Different temporal segments of the tones are indicated in the figure.

T h e attack time is sometimes also called the 'rise time', and its definition
varies slightly depending on the author. An often-used definition is the time
interval between the point the audio signal reaches 20% of its m a x i m u m value
178 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

and the point it reaches 80% of its maximum value [511]. Sometimes the
logarithm of the attack time is used instead of the raw value:

LAT = logio(^80-^2o), (6.17)

where t2o and tgo denote the beginning and end of the attack, respectively.
The temporal centroid measures the balancing point of the amplitude en-
velope of a sound, and it is calculated as

where E{n) denotes the RMS level of the sound at time n, and the summation
extends over a fixed-length segment starting at the onset of the sound.
The term vibrato refers to a periodic oscillation of the fundamental fre-
quency of a sound. Vibrato has proven to be quite a useful feature for instru-
ment discrimination, whereas this does not seem to be the case with tremolo^
which refers to a periodic oscillation in amplitude. Vibrato is characteristic
for string instruments, reeds, and the human singing voice. Vibrato can be
described by its rate, which is usually between 4 and 8 Hz, and its depth,
which is usually less than one semitone. Techniques for estimating the rate
and depth of vibrato are described in Chapter 12, Section 12.4.4.

6.3.5 Feature Transformation

Transforming features can enhance classification performance because it helps

to reshape feature probability distributions or reveal interesting structures
and, consequently, class boundaries can be more separable. Scale transform
and feature projection are two strategies to change the features in order to
get a 'better' feature set, compared to the original one.

Scale Transformation

One of the goals of features transformation is to improve their Gaussian-

ity. Many pattern recognition techniques such as linear discriminant analysis
(see Chapter 2) are optimal whenever the features have been sampled from
Gaussian distributions. A significant deviation from Gaussianity can increase
the classifier error rate. Osborne [490] discusses several transforms that change
the features scale non-linearly and can help in making them more Gaussian:
Square root, that is, substituting the original features by their square roots;
Logarithm, that is, taking natural logarithm of the original features, and
sometimes adding 1 before that in order to avoid having to compute log-
arithms of 0;
Inverse, that is, substituting 1/x for x, where x is the original feature;
 6 Automatic Classification of Pitched Musical Instrument Sounds 179

Arcsine-root [410], which consists of taking the arcsine of the square root
of the initial feature value; this transform spreads the feature values to the
tails of the distribution (that is, preferably away from its central part).
Therefore, this transform is indicated when the features are proportions
such as the ratio of band energy to the total spectral energy.
Apart from these simple transforms, the Box-Cox power transform [47] is
a standard tool for increasing the features Gaussianity, hence it is especially
recommended when working with Gaussian mixture classifiers [510]. It can be
computed as

where the value of A is chosen in order to maximize the Gaussianity of the

feature distribution.

Projection

Projections aim at reducing the dimensionality of the feature vectors by keep-

ing a reduced set of dimensions. In other words, given a feature vector x
composed of several features, projections only retain a subset of the initial
features (or linear combinations of them), resulting in a lower-dimensional
feature vector. When applied in the lower-dimension feature space, the classi-
fication algorithms are expected to have improved performance. In addition to
multidimensional scaling and canonical discriminant analysis, which are pre-
sented below, other techniques are principal component analysis (see Chap-
ter 2), independent components analysis (see Chapter 9 and [304]), projection
pursuit [197], and random projection [39].

Multidimensional Scaling

The motivation behind multidimensional scaling (MDS) is to produce a space

with typically 2 or 3 dimensions where projected, lower-dimensional feature
vectors are arranged topographically so as to yield a visual represention of
the similarity or dissimilarity of these feature vectors. Borg and Groenen [45]
provide a thorough presentation of this technique.
Multidimensional scaling first requires the definition of similarities, or
proximities, pij between any two objectshere between two sounds i and
j (i,j = l , . . . , m ) . These similarities may be obtained by numerical com-
putations from the sounds (via the features), or by subjective evaluation by
listeners. Given these proximities, MDS tries to represent the objects as points
in a Euclidean space Af with low dimension d^ such that a large distance be-
tween any two points in A' represents a small proximity between the sounds,
and conversely. Of course, this representation is not obtained by an exact
mapping of the proximities to distances, which generally do not exist since
the initial space is larger dimensional than A'. In order to implement MDS, we
180 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

need to select the distance measure to be used in X to represent the inverse

proximities; usually one selects the Euclidean distance between two objects
o^ and Oj in X:

{
dx \ ^^^

f;[o,(/)-o,(0]'i . (6.20)
Given these distances, we need to learn a function h that approximately
maps the proximities pij to the distances d(oi,Oj), and the location (the
coordinates) of the objects Oj, z = 1 , . . . , m in A*. This is done by minimizing
the stress function
r n n -i 1/2

i=l j>i
S(h,oi,. . . , o ^ (6.21)

2=1 j>i

The minimization is performed numerically with respect to h and to the

objects o^, i = 1 , . . . ,m, by moving them around iteratively until the stress
is minimum. The stress function gives the relative error that, on average,
the distances d(oi,Oj) in X differ from the mapped proximities h{pij). When
using the Euclidean distance d(-,-), the final coordinates of the o^'s in X
do represent the principal coordinates which would be obtained when doing
PC A on the dissimilarity matrix [h(pij)]. ALSCAL and PROXSCAL are two
iterative algorithms designed to minimize stress when the assumed distances
are not Euclidean.

Canonical Discriminant Analysis

Canonical discriminant analysis (CDA) is a dimension reduction technique re-

lated to principal component analysis and canonical correlation. In a canon-
ical discriminant analysis, we find linear combinations of the features that
provide maximal separation between the classes or groups. CD A derives a
set of canonical variables that are linear combinations of the features, and
that summarize between-class variation in much the same way that principal
component analysis summarizes total variation; see Chapter 2. Canonical dis-
criminants are computed iteratively: the first canonical discriminant function
is the best single linear combination of attributes (that is, dimensions in the
feature vectors) that discriminates between the classes. The second canonical
discriminant function is the best single linear combination orthogonal to the
first, and so on. Similar to principal component analysis, each canonical dis-
criminant function is associated to an eigenvalue which indicates its relative
discriminating power.
 6 Automatic Classification of Pitched Musical Instrument Sounds 181

Canonical discriminants being linear combinations of the original features,

they are projections, which are computed as follows. First, the between-class
scatter matrix SB and the within-class scatter matrix Sw are computed as

1 ^
i:B = 7 ^ m , ( M j - M ) ( M ^ - / i r , (6-22)

1 ^

j=l x|y=j

where J is the number of classes, /x is the mean of all available m observations,

fij is the mean of the observations in class j , y G { 1 , 2 , . . . , J } denotes the
class label of x, ruj is the number of observations in class j , and x is a
given feature. Then, the optimal projection W is the projection matrix which
maximizes the ratio of the determinant of the between-class scatter to the
determinant of the within-class scatter of the projections (also called Fisher's
criterion):
|w"ri:BW| , , ,^^,,
^ = " ^ ^ r " IWT^wWl ^ [ w w . . . , w , J , (6.24)
where {w^|2 == 1,2, ...,(ix} is the set of generalized eigenvectors of UB
and X'w, corresponding to the dx largest generalized eigenvalues {Xi\i =
l , 2 , . . . , d x } [295]. The matrix W can be found by solving the generalized
eigenvalue problem X'BA* = AZ'wA* by means of simultaneous diagonal-
ization [202]. To avoid singularities, especially if the features are sparse, one
can apply principal component analysis first to reduce the dimension of the
feature space to dx J, and then use CDA to reduce the dimension to J 1.

6.3.6 Feature Selection

Using very large sets of features for building an automatic classification sys-
tem is usually to be discouraged: First, some features can be redundant or
irrelevant; second, the computational cost for using many of them might be
high; and third, some features can be misleading or inconsistent regarding
the task, and consequently the classification errors may increase. In any case,
interpreting a model containing a large set of features can be very difficult or
even impossible. In general, the informal recommendation is to use ten times
fewer features than training instances^ [313]. Selecting features can be done
on a ranking basis (i.e., evaluating one feature after another) or on a best-set
basis (i.e., evaluating subsets of features in a global way) [262].
We list below three different strategies in order to find a near-optimal
number of features for a classification task [42]:
Recent techniques such as support vector machines (see Chapter 2), however,
are less subject to dimensionality concerns. In any case, including misleading features
lowers the performance.
182 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

Embedding makes the feature selection stage intertwined with the classifi-
cation algorithm, as in the case with decision trees, or with discriminant
analysis.
Filtering decouples feature selection from the model learning process by
first applying a feature selection over the original feature set, and then
feeding the classification algorithm with the selected features only.
Wrapping uses a features evaluation step which is intimately connected
with the learning process: it uses the prediction performance of a given
learning algorithm to assess the relative usefulness of subsets of features.
Theoretically, this strategy should be the best one [42], but the price paid
is a high computation time (this is an NP-hard problem).
In addition to selecting the features with respect to a classification algo-
rithm, we must decide on an evaluation criterion. The information gain is the
standard criterion used to build decision trees [467], but it can be also used
to rank the importance of features, outside of the decision trees framework.
In order to characterize the amount of information carried by a feature (e.g.,
the zero crossing rate), we study its influence on the entropy of the full set
of features, via the information gain. Let x denote the vector made of several
examples of a given feature (e.g., the zero crossing rate over several frames)
and let X be the set of all the features, each of which being extracted over sev-
eral frames. The entropy H(a) of a set of random variables a with probability
density function p(a) is defined as^

H(a) = - y p ( a ) l o g p ( a ) d a , (6.25)

when a is a continuous random variable; the discrete case is obtained by

replacing the integral by a sum over all possible values of X. We define the
information gain of the feature x relative to the set of features X as

5(X,x) = H(X)-H(X|x), (6.26)

where H(X|x) is the conditional entropy of X when the value of x is known. In

other words, the information gain measures the amount of information added
by considering the full feature set X with entropy H(X), as opposed to the
reduced set X\x with entropy H(X|x).
The correlation-based feature selection technique of Hall [268] has pro-
vided very good results in selecting features for automatic classification of
sounds [288], [510]. We define the following merit heuristic of a feature subset
Xs containing m features, each having several values extracted from various
frames:
M(X3) ^ , ^^;^'^ ^^ , (6.27)

^Random variables and probability density functions are introduced in Chapter 2.

 6 Automatic Classification of Pitched Musical Instrument Sounds 183

where Cfc is the average feature-class correlation and Cff is the average feature-
feature intercorrelation.^ We can interpret the numerator of (6.27) as an in-
dicator of how a set of features is representative of the class, whereas the
denominator indicates how much redundancy there is among the features.
The correlation-based feature selection technique is not a ranking method,
as it works with subsets of features. In practice, subsets are examined using
either backward search or forward search of best first search. Backward search
consists of starting from the full feature set and greedily removing one fea-
ture at a time as long as the evaluation does not degrade too much). Forward
search starts from an empty set of features and greedily adds one feature at
a time until no possible single feature addition results in a higher evaluation.
Best first search starts with either no features or all the features and examines
a given number of consecutive expansions or reductions of the existing subset
in order to find local improvements over the current best subset [540], [267].
In the context of classification, given a training set of features grouped
in J classes, Peeters [510] has proposed audio feature selection using inertia
ratio maximization using feature space projection (IRMFSP), which seems to
compare advantageously with other effective algorithms and has also been
used by other researchers in sound classification [416], [180]. IRMFSP selects
first the best features according to the value R[i] for the feature # i (where the
index i refers to a given type of feature, e.g., the zero crossing rate) defined
as

E fnj\\^lj[i] - ix[i]f
m = N '^r , (6-28)
Eii^"i ^]-^l[i]f
n=l

where Jj is the number of training features in class j ( j = l , . . . ,J), /Li[i] is the

empirical average of feature # i in the full training set, /x^[i] is the empirical
average of the feature #2 for features in the training set belonging to class j .
The ratio in (6.28) is the between-class inertia to the total inertia: the larger
R[2], the more discriminative is the feature. The feature selection proceeds as
follows. First, the features are selected from the largest inertia to the small-
est inertia. This selection process is intertwined with an orthogonalization of
the feature space; that is, each new selected feature is made othogonal to
the previous ones (this is the standard Gram-Schmidt orthogonalization; see
e.g., [388]).
The extractor discovery system (EDS) [692] is an interesting tool that
'creates' new features by means of genetic programming [372] applied to

^The term 'correlation' was used by Hall in its general sense, without referring
specifically to the classical correlation [267, p. 51]. However, in his implementation
in the free software Weka [674] (www.cs.waikato.ac.nz/ml/weka), the author used
the classical variance-normalized correlation, removing, however, the mean of the
data before calculating the correlation.
184 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

the existing features and to a library of mathematical and signal processing

functions. When used in instrument classification problems (percussion/no-
percussion detection and 4-class string instruments discrimination), the EDS
increased between 1 percent-units and 9 percent-units the performance com-
pared to using the original 'simpler' features [691]. A somewhat similar system
has also been presented by Mierswa and Morik [463].
Other interesting approaches to feature selection, based on genetic algo-
rithms (GA) are by Fujinaga et al. [200], [199] and about rough sets [359],
[360], [367], [669], [671]. The latter have been used to select features for in-
strument classifiers. Further details about genetic algorithms and rough sets
may be found, respectively, in [217] and [508].

6.4 Classification Techniques

When designing a classifier for musical instrument sounds, researchers are

usually faced with the problem of selecting a single classifier for all the instru-
ment sound classes or, conversely, using several 'focused' classifiers. In the first
case classifiers are termed flat classifiers, whereas the second case classifiers
may be one of several diS'erent solutions ranging from hierarchical classifiers
to ensembles of classifiers.
In the hierarchical case [443], [176], [515], a classifier is used in each of the
different decision nodes of a hierarchy that has been created by means of
knowledge-based decision rules. A hierarchical classifier may decide first if
the sound belongs to a given broad class (such as impulsive vs. continuous).
Then, the sound is classified into a finer class that is conditioned by the
first classification decision (for instance, in the case of a continuous sound,
the second level of the hierarchy decides if it is from woodwinds, bowed
strings or brass; in the case of an impulsive sound, the options could be
plucked strings or percussion). The final level of the hierarchy, depending
on the previous decisions, assigns the specific instrument class (guitar,
oboe, etc.).
Concerning ensembles of classifiers, different classification techniques may
be used in parallel or serially, in order to deal with subsets of the original
data, of the feature set, or with the whole set; see [379].
Some of the classification techniques that have been frequently used for the
automatic classification of musical instrument sounds are explained in Chap-
ter 2. This is the case with support vector machines (SVMs), and Gaussian
mixture models (GMMs), which are therefore not included here. In this sec-
tion, we discuss instance-based classifiers, discriminant analysis, decision trees,
and artificial neural networks.
 6 Automatic Classification of Pitched Musical Instrument Sounds 185

6.4.1 Instance-Based Learning

Instance-based learning or lazy learning refers to the idea of avoiding the

computation of any abstract model for a class of objects. Instead, techniques
under this category assume that instances that are closer to the one we try to
classify (i.e., they are 'similar' in the feature space) will provide the required
class label. Instance-based learning is usually the most workable option when
the number of available classes is extremely large, or cannot be totally antici-
pated, as it is the case, for instance, in the domain of classification of sound
effects [62].
The k-nearest neighbours (/c-NN) algorithm is one of the most popular
algorithms for instance-based learning. It first stores the feature vectors of
all the training examples and then, for classifying a new instance, it finds
a set of k nearest training examples in the feature space, and assigns the
new example to the class that has the largest number of examples in the set
(see Fig. 6.5). Traditionally, the EucHdean distance measure (6.20) is used to
determine similarity, and the number of neighbours is determined empirically;
see the description of /c-NN in Chapter 2. Although the /c-NN technique is easy
to implement, it has some drawbacks [467]:
It requires having all the training instances in memory in order to yield a
decision for classifying a new instance;
It may require a significant computational load each time a new query is
processed;
It is highly sensitive to irrelevant features that can dominate the distance
metrics;
It does not provide a generalization mechanism (because it is only based on
local information), although several techniques (K-means clustering, linear
vector quantization) can be used to compute 'prototypes' from several
nearest neighbours.
The fc-NN algorithm has probably been the most used in classification
studies of musical instrument sounds, achieving very high rates of performance
under most of the tested conditions; see [200], [442], [11], [174], [416], [325].
Given its simplicity, it is usual to include it as a kind of 'reference' when
comparing different classification algorithms.

6.4.2 Discriminant Analysis

Discriminant analysis (DA) includes several variants of the generic idea of de-
riving a discrimination function (i.e., one that separates two classes of objects)
that is a weighted combination of a subset of the features used to character-
ize a series of observations. As we have seen in previous sections, this idea
can be also used for selecting and projecting features by minimizing the ra-
tio of within-class scatter to the between-class scatter. A thorough formal
186 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

2 L
h B /
I A\m

I -.. A..--
r A

Fig. 6.5. An illustration of /c-NN classification. The point marked with a star would
be classified as belonging to category B when /c = 3 (as 2 out of its 3 neighbours are
from class B; but in case of using k = 5 classification would be A because there are
3 nearest neighbours belonging to this category and only 2 belonging to B.

and practical treatment of this techniques can be found in the monograph by

McLachlan [453].
Linear discriminant analysis (LDA) is a minimum-distance classification
method that uses the empirical mean and covariance matrix of the training set
classes. The data are projected so as to eliminate inter-feature correlations and
to standardize the variance of each variable, after which Euclidean distances
are computed in the standardized space; see Chapter 2 for a presentation. LDA
assumes that the classes have a common covariance matrix Ci but sometimes
this cannot be true. In that case the quadratic discriminant analysis (QDA)
can be applied, where the decision boundary between each pair of classes is
described by a quadratic equation. These and other variations such as logistic
or regularized discriminant analysis have been been used in several studies on
classification of musical instruments [443], [11], [12], [510], [416], [552].

6.4.3 Decision Trees

Decision trees are pervasively used for different machine learning and classifi-
cation tasks. One of the main reasons for their popularity may lie in the fact
that they produce a simple classification procedure which can be interpreted
and understood as a series of 'if-then' actions. Decision trees are constructed
top-down, beginning with the feature that seems to be the most informative.
 6 Automatic Classification of Pitched Musical Instrument Sounds 187

that is, the one that maximally reduces entropy according to the information
gain measure (6.26). For this feature, several branches are created: one for each
of its possible (discrete) values. In the case of non-discrete valued features,
a procedure for discretization of the value range must be defined (see, for
instance, [183]). The training data are assigned to a descendant node^ at the
bottom of the branch that corresponds their value. This process is repeated
recursively starting from each descendant node. An in-depth treatment of de-
cision trees can be found in Mitchell [467] and Duda et al. [161]. Quinlan's IDS
and C4'5 [533] are among the most popular algorithms for building decision
trees.
Decision trees used for the classification of instrument sounds have usually
yielded worse results than other classification methods [288], [510]. Otherwise
they have provided hints on the nature of the features and values that dis-
criminate among pitched instrument classes [319], [668]. Recent enhancements
to basic decision trees such as AdaBoost [196] or Random Forests [51] may
provide results that are competitive with other cutting-edge classification al-
gorithms.

6.4.4 Artificial Neural Networks

An artificial neural network (ANN) is an information-processing structure

that is composed of a large number of highly interconnected elementscalled
neurons or unitsworking in unison to solve a specific problem. Neurons are
grouped into layers (which can be input, hidden, and output) to be intercon-
nected through different patterns (see Fig. 6.6). Each neuron has an activation
function. The activation of each neuron depends on the activity coming from
the other neurons that connect to it, and on the activation function. The full
neural network learns a complex, non-linear function between input (here, fea-
tures extracted from a given sound) and output vectors (here, an instrument
class) by changing the interconnections between the neurons: a training set
containing example input and related output vectors is presented to the ANN,
which learns the non-linear function. This is a regularization problem, as de-
scribed on p. 57. The interested reader is referred to the specific monograph
of Bishop [40] for an in-depth presentation of the subject.
One of the most popular neural network models is the multilayer perceptron
(MLP). Usually, the action of each neuron is modelled as a linear function of its
inputs, and the activation is modelled by a sigmoid function [279]. Structure-
wise, MLP is a spatially iterative neural network with several layers of hidden
neuron units between the input and output neuron layers. The most commonly
used learning scheme for the MLP is the back-propagation algorithm [555].
The weight updating for the hidden layers adopts the mechanism of back-
propagated corrective signal from the output layer. It was demonstrated that
perceptrons with one hidden layer are adequate as universal approximators of
any non-linear function [666], which means that they can theoretically learn
any non-linear relation between their inputs and outputs. This makes MLPs
188 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

Input units

Hidden units

Output units

Fig. 6.6. A diagram of an artificial neural network with 8 input units, a hidden
layer with 3 units, and an output layer with 8 units.

a good choice when the function to be learned is not known in advance, or it

is suspected to be non-linear.
ANNs do have some disadvantages, though: the computation time for
the learning phase may be very long, adjustment of user-defined parameters
can be tedious and prohibitively time consuming, and data over-fitting can
degrade their generalization capabilities. Interest in ANNs appears to have
declined since the arrival of SVMs. This can be explained because of several
advantages of the latter over ANNs: SVMs require fewer parameters to be
tuned and less time to be trained/used, and are able to provide similar or
greater accuracy than ANNs, with minimal over-fitting. However, when ei-
ther ANNs or SVMs are learned, the classification decision is very fast when
compared to other popular methods such as /c-NN, and they can also learn to
disregard irrelevant attributes.
Kostek [360] and Park [495] are probably the most exhaustive studies on
automatic instrument classification using neural networks.

6.5 Classification of Isolated Sounds

In this section, we review research aimed at developing systems for the auto-
matic classification of isolated sounds of pitched musical instruments. As we
will see, for the more complex tasks of recognizing instruments in solo and
duet phrases or in polyphonic music, the fact that each research team has
been using a different test database makes it unfair and unreliable to make
direct comparisons between the reported classification accuracies. The goal of
this section is, hence, to provide enough information so that the reader can
understand the work done and to evaluate the main outcomes of the research
himself. We first examine flat systems, which provide instrument labels in a
single classification step. Then we will move to systems that, instead of di-
rectly deciding the instrumental class, proceed from broader categories to the
more specific ones, following a hierarchical classification scheme.
 6 Automatic Classification of Pitched Musical Instrument Sounds 189

6.5.1 Flat Classification Systems

An important distinction to be made first is that between systems that have

been developed and tested using several different databasesmeaning that
several different 'instances' of each instrument (for example, several pianos),
different recording conditions, and players are consideredand those using
only one database (mostly MUMS; see Section 6.2.1). The systems in the
latter group are only learning to identify a given instrument instance, for
example a particular violin, and therefore their generalization capabilities are
suspect. Martin [442] and Eronen [174] observed this fact indirectly, and a
conclusive demonstration was provided by Livshin and Rodet in [414] (but
also see [285]), where a classifier trained and tested using the same large
collection (IRCAM's Studio On-Line) achieved excellent performance rates
(96% for instrument classes), but when tested using a different collection of
sounds, performance dropped down to 26%.
A summary of classification systems for isolated instrument sounds is pre-
sented in Table 6.3. It should be noted that only systems dealing with multi-
ple instances of each instrument are included in the table. Also, definitions of
the acoustic features may vary slightly from the formulas given in Section 6.3,
depending on the authors. The data shown is not repeated below, but instead,
some general remarks are made in the following.
Although most of the reviewed systems based their features on a fast
Fourier transform (FFT) front-end, some studies have reported results on
wavelets [371], correlograms (see Chapter 8, p. 245) [442], Hough transforms
[552], linear prediction [174], [682], [87], and constant Q transforms [56], [57],
[325]. An advantage of FFT-based features over wavelet-based features has
been reported in [361], and an advantage of linear prediction over cepstral
analysis has been reported in [375]. It is clear that more comparative research
is needed to make strong conclusions regarding the time-frequency analysis
front-ends.
Fundamental frequency (FO) estimation is an important step as this fea-
ture by itself, and harmonic descriptors based on it, may provide substantial
discriminative information. The dependence of timbre features on FO makes
it difficult to compare the feature values of two instruments playing different
notes. Kitahara et al. [345] developed a musical instrument classification sys-
tem using an FO-dependent multivariate normal distribution, where the mean
of the distribution of each feature was represented as a cubic polynomial
function of the fundamental frequency This FO-dependent mean function was
used to represent the pitch dependency of each feature, while FO-normalized
covariances after substracting the pitch contribution were used to represent
the non-pitch dependence of the features. Musical instrument sounds were first
analysed by the FO-dependent multivariate distribution, and then identified
by using a discriminant function based on the Bayes rule. A slight advantage
of considering the FO-dependence was observed both at the instrument-level
190 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

Table 6.3. Summary of selected research on automatic classification of isolated

instrument sounds. Only studies dealing with several different instrument instances
and recording conditions have been included. The column NC shows the number
of classes in each system. In the last two columns, the performance of hierarchical
systems has been indicated with '(H: )'.

Author, Total instances'^ NC Acoustic features Classifi- Instru- Family

year [ref] cation ment perfor-
algo- perfor- mance
rithm mance (%)
(%)
Eronen, 5286 (MUMS, Iowa, 29 M F C C ( a t t a c k - s t e a d y ) , FO, k-NN 35 77
2001 [174] SOL, R o l a n d X P 3 0 , A T T , onset features, SC, (H: 30) (H: 75)
own recordings) Crest Factor, AM

Livshin 4381 (SOL, Iowa, 16 SC, A T T , t e m p o r a l LDA &: 47-69 62-92

et al., 2003 M U M S , Prosonus, decrease, T R I , HD, S K W , k-NN
[413] Vitous) KUR, SV, SS, M F C C ,
noisiness

Peeters, 4163 (SOL, Iowa, 23 same as above LDA & 54 81

2003 [511] M U M S , Microsoft GMM (H: 64) (H: 85)
MI, Prosonus, (hierar-
Vitous) chical)

Eronen, 5895 (MUMS, Iowa, 7 M F C C , d e l t a - M F C C -\- ICA H M M 68 n/a

2003 [175] SOL, M a r t i n , own
recordings)

Kitahara 6247 (RWC) 19 SC, O E R , FO relative Bayes 80 91

et al., 2003 energy, KUR, S K W , F M , (fc-NN
[345] a m p l i t u d e envelope slope, after
onset energy PCA &
LDA)

Kostek n / a (CMIS, MUMS) 12 Wavelet-based energy MLP 71 n/a

et al., 2004 bands, M P E G - 7 features
[362]

Szczuko 2517 (CMIS, 16 M P E G - 7 features, O E R , FO M L P 86 n/a

et al., 2004 MUMS) (2-stage (H: 89)
[615] hierarchi-
cal
MLP)

P a r k et al., 829 (several 12 SS, SC, harmonic slope, MLP 71 88

2005 [496] commercial L P C noise, harmonic with el-
instrument-sample expansion/contraction, liptical/
CDs) spectral j i t t e r and shimmer, radial
spectral flux, T C , ZCR basis
functions

Chetry 4415 (Iowa, RWC, 11 Line s p e c t r u m frequencies K-means 95 n/a

et al., 2005 voice) derived
|[88] codebook

In t h e Total instances column, t h e a b b r e v i a t e d sound collections are McGill University

Master Samples ( M U M S ) , University of Iowa samples ('Iowa'), I R C A M ' s Studio On-Line (SOL),
G d a n s k ' s Catalogue of Musical I n s t r u m e n t Sound (CMIS), and Real World C o m p u t i n g (RWC).
O t h e r n a m e s refer t o commercial sample C D - R O M s .
In t h e Acoustic features column, SC denotes spectral centroid, SS is spectral spread, SV
is spectral variation, S K W is skewness, K U R is kurtosis, SFM is spectral flatness, Z C R is zero
crossing r a t e , T R I is tristimulus, O E R is odd-to-even ratio, HD is harmonic deviation, A T T is
a t t a c k t i m e , T C is t e m p o r a l centroid, F M denotes v i b r a t o , and AM is a m p l i t u d e m o d u l a t i o n .
 6 Automatic Classification of Pitched Musical Instrument Sounds 191

and at the family-level classification. A different approach to this problem can

also be found in [688].
Evidence abounds on the positive effect of combining different types of
features. This has been demonstrated for wavelet-based energy bands plus
MPEG-7 audio features [362], for spectral plus temporal features [670], and
for MFCC plus other spectral and temporal features [412], [510], [597].
Feature-vector dimensionality reduction by projection is another recom-
mendable option, as the work of Peeters et al. [515], [511], Livshin et al.
[413], [416], Kitahara et al. [345], and Essid et al. [177] illustrates. In all these
cases, sets with hundreds of features were reduced by means of discriminant
analysis (see Section 6.3.5 above) to a fraction of the original dimensionality
without decreasing, or even effectively increasing, the discrimination rate of
the system. Another interesting projection approach has been proposed by
Kaminskyj and Czaszejko [325], who applied PCA to a vector consisting of
the constant-Q transform frequency bins in the vicinity of the first 20 har-
monic overtones. The PCA-transformed feature vectors in successive analysis
windows formed a multidimensional abstract trajectory space, where sounds
of different classes showed different trajectories.
Regarding the classification techniques, /c-NN is the most-often used and
one of the best performers in the this kind of problem. Being very easy to
implement, it is a good point of comparison for other classifiers. LDA is some-
times suggested as another 'default' method to be included in comparative
studies but unless prior feature selection and projection is properly done, it
cannot compete with SVMs or GMMs. We expect to see an increasing number
of studies using SVMs, as they have theoretical and practical advantages that
are well understood and clearly demonstrated in other areas of data classifi-
cation (see Chapter 2).
Artificial neural networks have been used almost exclusively by Kostek
et al. [357], [358], [368], [360], [366], [363], [364], who also reported on perform-
ing instrument classification using rough sets. The latter, and other rule-based
techniques such as decision trees, have not proven to be competitive with other
classification techniques [597], [515], although some authors have found other
reasons for adopting them [361], [671]. The ANNs used by Kostek's team were
standard MLPs with one hidden layer. Explorations of parallel networks can
be found in [615], and networks with radial and elliptical basis functions have
been studied by Park et al. [495], [496]. Also, Park proposed a technique called
nearest centroid error clustering, which he used together with the standard
back-propagation learning algorithm for training the network. With this new
technique, the system achieved 71% classification accuracy, 25%-units more
than without the clustering.
A surprising observation is that hidden Markov models (HMMs) have not
been very usual in automatic instrument classification systems, given their
prevalence in speech recognition and speaker recognition (see Chapter 2).
Casey [71] provided an interesting classification framework, in connection
to MPEG-7 spectral bases, but the most advanced HMM-based system for
192 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

classifying pitched instrument sounds seems to be that of Eronen [175]. He

proposed to use independent component Analysis for transforming the origi-
nal feature vectors (which improved the performance by 6%) and to use a
discriminative training algorithm for the HMM, which also slightly improved
the performance.
To conclude, we have observed that the difficulty of classification is directly
related to the number of classes [442], [12]. We should then expect very good
performance for systems dealing with less than 10 classes. On the other hand,
working with more than 20 classes becomes an extremely difficult problem, not
only in the case of instrument sounds but also in any other automatic learning
domain. Hierarchical classifiers have been proposed as a means of maintaining
the generalization ability of a classifier when the number of classes is increased
[442]. These are discussed in the next section.

6.5.2 Classification Based on Hierarchical Decisions

One of the influential ideas of Martin [441], [442], [443] was to use a hierarchical
procedure consisting of (1) initial discrimination between pizzicato (plucked)
and sustained sounds, (2) discrimination between different instrument families
(e.g., strings, woodwind, and brass), and (3) depending on the previous deci-
sions, final classification into specific instrument categories. Other hierarchical
systems have been developed since then by Eronen et al. [176], [174], Agostini
et al. [11], Szczuko et al. [615] and finally, Peeters et al. [515], [510], which
is probably a fair representative of the current state of the art in instrument
classification.
Table 6.3 summarizes some facts about the systems mentioned in this
section. As can be seen, the results using a hierarchical instead of a flat clas-
siflcation scheme have not been conclusive: some authors report moderate
improvements, whereas others report moderate deterioration in error rates.
On the other hand, what is consistent is a trend of increasing difficulty when
the categorization of an instrument goes from the most abstract to the most
specific: the pizzicato versus sustained decision is very easy, the family classi-
fication problem is a bit more difficult, and finally, the specific assignment of
instrument labels still leaves some room for improvement.
Although the use of hierarchies is conceptually appealing, it is worth noting
that errors at each level are 'carried over' in a multiplicative way. Therefore, if
there is an error at the topmost level, it will be propagated to the lower levels
of the hierarchy. One rule of thumb for trying a hierarchical system would be to
look at the between-class confusion matrices and search for a large number of
confusions between instruments from different families. In this case, provided
a very efficient family discriminator, these confusions could be reduced by
means of a series of hierarchical decisions.
Instead of embedding some taxonomic knowledge into the classifier (i.e.
hardwiring the family and subfamily taxonomy), Agostini [10] and Kitahara
et al. [346] have approached the automatic building of instrument taxonomies
 6 Automatic Classification of Pitched Musical Instrument Sounds 193

by means of clustering techniques, although the reported results were worse

than those using a predefined taxonomy. A more elaborate proposal was
presented by Essid et al. [179], but the authors did not use their induced
taxonomy for classification purposes.
A related and promising thread for future explorations is an approach
called 'mixtures of experts' [593], where a sound is analysed by a number of
specialized classifiers, and the results are then combined by a voting mech-
anism which attempts to neutralize individual errors made by one of the
classifiers. For example, if a sound comes from a bowed violin, but the most
generic classifier erroneously classifies it as 'not-sustained', while the bowed-
string and violin experts assign it to the correct classes, then the sound would
be correctly labelled provided a conflict resolution mechanism is properly im-
plemented. FanelH et al. [181] presented a modular neuro-fuzzy network for
classifying instrument families, where different submodules operated on dis-
tinct subsets of input features. Similarly, Kostek et al. [371] presented a system
of parallel ANNs, where each network generated a decision based on a single
energy band and then all decisions were combined to produce a coherent out-
put. This idea of combining specially trained subsystems has not been pursued
so far in any other of the studies we have reviewed here, but it deserves more
attention (see also Essid et al. [177] in the next section).

6.6 Classification of Sounds from Music Files

One of the research trends in the classification of musical sounds is moving
toward more ecological scenarios, that is, working with real music files, ad-
dressing practical applications, and utilizing musicological information in the
process. Three different classification contexts, in the order of increasing com-
plexity, are reviewed here: solo phrases, duets, and complex musical mixtures.

6.6.1 Classification of Instruments in Solo Phrases

Detecting the instrument played in a solo phrase should in principle be feasible
by applying the same algorithms and models that are used when working
with isolated notes. Even though we know that articulation and expressive
phrasing elements may convey cues for instrument identification [340], the
existing research has not exploited these yet. Research has also disregarded,
up until now, the problem of how to locate the solo sections in an entire
musical piece, and all the works reviewed here process solo phrases that have
been manually extracted.
Table 6.4 summarizes some studies dealing with this problem. The target
material, features, and other details of the systems are not repeated here;
instead we highlight other aspects that also deserve attention.
One important decision in solo instrument identification is how long it
takes to decide on the instrument label. Using longer time segments and cal-
culating averages or choosing the most frequent class decision should provide
194 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

Table 6.4. Summary of research on instrument classification from solo phrases.

The column NC shows the number of classes in each system. See the footnotes
below Table 6.3 for interpreting the abbreviated feature names and sound collections.
Author, Acoustic material NC Acoustic features Classifi- Instru-
year [Ref] cation ment
algo- Perfor-
rithm mance
(%)
Martin, 1500 excerpts (MUMS, 14 A T T , SC, SI, FM, AM, Hierarchi- 39
1999 [442] CDs, own recordings) ZCR, onset features cal
k-NN
Marques &: 1638 (train) + 171 (test) 8 MFCC SVM 70
Moreno, seconds from 100 music
1999 [437] tracks on CDs
Brown 14 commercial CDs 4 Cepstral coefficients, GMM 80
et al., 2001 autocorrelation, SC,
[57] constant-Q band
differences
Ventura- 1800 (train) + 900 (test) 6 P e r c e p t u a l linear HMM 94
Miravet seconds per class from CDs prediction,
et al., 2003 (7 different players per Linear-prediction derived
[640] class) cepstral coefficients
Livshin 108 solos (commercial CDs) 7 SC, KUR, SKW, SS, LDA + 88 (85
et al., 2003 slope, T R I , M F C C k-NN real-
[416] time)
Vincent & Train: sounds from RWC; 5 non-linear ISA GMM 90
Rodet, test: 2 x 5 seconds excerpts
2004 [647] from 10 solo recordings per
class
Essid 900 (train) + 90 (test) 5 MFCC GMM 67
et al., 2004 seconds per class from
[177] M U M S , Iowa, SOL, and
commercial music CDs
Essid 900 (train) + 150 (test) 10 M F C C , d e l t a - M F C C , SC, SVM 77
et al., 2004 seconds per class from same SS, S K W , KUR, bandwise
[178] as above ( + i n s t r u m e n t s SFM
added)
K r i s h n a &; Train: Iowa; test: RWC 3 Line S p e c t r u m GMM 74
Sreenivas, Frequencies, M F C C , L P C
2004 [375]
1 Chetry 300 (train) + 300 (test) 6 Line S p e c t r u m Frequencies SVM 78
et al., 2005 seconds per class from CDs plus their velocity and
|[87] acceleration

better results, even leading to better discrimination rates t h a n those in classi-

fying isolated sounds. The optimal reported durations vary between two and
ten seconds, depending on the authors [437], [640], [178]. As the time dimen-
sion has a more prominent role here t h a n in identifying isolated sounds, HMMs
may have some advantage over 'static' modelling techniques such as those used
in t h e previous section. T h e HMM-based system reported by Ventura-Miravet
et al. [640] provided impressive evidence on t h a t .
An original proposal t h a t goes back to the idea of 'mixture of experts' is
using a 'one against one' classification strategy, where a series of classifiers
are devised t h a t only make binary decisions between an instrument A and
 6 Automatic Classification of Pitched Musical Instrument Sounds 195

instrument B, and the decisions are then combined by voting [177]. In addition
to improving the classification accuracy under certain circumstances, the use
of pair-wise classifiers makes it possible to find some characteristic features
that are very useful in a given pair-wise discrimination but remain useless in
others.
Exploring new features and increasing their number is another trend ob-
served in this research context. The log-energy within octave sub-bands and
the logarithm of the energy ratio of adjacent sub-bands have been used as
a way to characterize the spectral energy distribution of a sound [177]. Line
spectral frequencies have also been incorporated as a way to model the reso-
nances and peaks of the power spectrum with higher precision and robustness
than linear prediction coefficients (LPC) [375], [87].
Independent subspace analysis (ISA) has been successfully used for per-
cussive sound classification (see Chapter 5) but it has only been tested for
classifying pitched sounds by Eronen [175] and by Vincent and Rodet [647].
The latter represented the short-time spectrum of musical excerpts as a non-
linear weighted sum of typical spectra plus noise, which were learned using
files from a database containing isolated notes and solo recordings. These tem-
plates were then used to determine, with a very high accuracy, the instrument
played in the solos of commercial recordings (even when they were artificially
distorted with reverberation or noise). The authors showed that their model
has some theoretical advantages over methods based on GMMs or on linear
ISA and that it worked successfully even for the classification of instruments
in duets.
A surprising observation is that any of the reviewed systems do not seg-
ment solo phrases into notes. The addition of a reliable onset detector would
allow the inclusion of envelope-related temporal features and the reduction
of computational load by doing the actual classification only once per note
onset. Onset detection is, however, a hard problem in the case of music signals
that do not contain percussion instruments [95].

6.6.2 Classification of Instruments in Duets

Classification of instruments in musical duets requires more complex tech-

niques than those for isolated samples or solo phrases. The added difficulty
lies in achieving some kind of sound source separation, or robust extraction
of the features of a given source in the interference of the other source. As we
will see, the few existing studies that have been summarized in Table 6.5 are
still more or less exploratory in nature: in addition to the sound separation
issue, they use few duet types and a small music collection, so their reported
high performance estimations should be taken with care.
Two different approaches have been examined here. In the first one, primi-
tive source separation (or source reduction, as Livshin and Rodet [416] termed
it) is performed by estimating the FO of one or both instruments and then
substracting the partials associated with the first sound in order to estimate
196 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

Table 6.5. Summary of research on automatic identification of instruments from

duets.
Author, Acoustic material Number Acoustic Separat- Classi- Performance
year of pairs features ion fier (%)
[Ref]
Livshin Training: isolated 8 (7 dif- SC, A T T , Source LDA 40-100, de-
& Rodet, sounds a n d sounds ferent temporal reduction projec- p e n d i n g on
2004 from solo phrases. instru- decrease, T R I S , (spectral tion -|- t h e pairs
[416] Test: 18 music pieces ments) HD, S K W , subtrac- k-NN
KUR, SV, SS, tion)
M F C C , noisiness
Kostek Training: 400 sounds, 8 (4 dif- P a r t i a l s ' Spectral MLP n/a (only
et al., plus 400 sounds for ferent amplitudes subtrac- average
2004 cross-validation of t h e instru- tion activation
[362] t i m b r e model. Test: ments) of t h e M L P
artificial pairs o u t p u t layer
reported)
Eggink Sz Training: isolated 10 (5 dif- Energy within Missing GMM 74
Brown, sounds mixed in pairs ferent 120 s u b - b a n d s feature
2004 and artificially mixed instru-
[167] solo phrases. Test: ments)
mixed pairs of sounds

the spectrum of the second one. Using this approach, harmonic cancellations
and erroneous enhancement of partials should be expected, depending on the
degree of overlap between the spectra of the two sounds.
Livshin and Rodet [416] approached the classification of instruments in
duets by using their real-time solo recognition system (see Table 6.4). In or-
der to detect instruments in duets, the system first estimated the two FOs
using an algorithm by Yeh and Robel [681], and also computed their corre-
sponding harmonic partials. The fundamental frequencies were quantized to
the nearest musical note, and contiguous frames with the same value were
chunked together. Each chunk was then used twice in a phase-vocoder fil-
tering process of source reduction: in the first pass, all the harmonics of the
estimated fundamental were kept, whereas in the second pass, the sustained
note's harmonics were filtered out and the 'residual' partials were kept. Over-
lapping harmonics of the two notes were not filtered out. Finally, the partials
of the fundamental (intended to correspond to one instrument) and the resid-
ual partials (intended to correspond to the other instrument) were sent to a
classifier in order to generate their corresponding labels.
Kostek et al. [365], [362] proposed the decomposition of duet sounds based
on the modified frequency envelope distribution (FED) analysis, which was
originally described in [370]. The FED algorithm decomposes a signal into a
linear expansions of sinusoids with time-varying amplitudes and phases. The
first step of the duet analysis method is the estimation of the FO of the lower-
pitched instrument. The input signal is divided into short overlapping blocks,
and FO is estimated for each block separately to deliver the FO contour. Using
the FO information and the FED algorithm, the time-varying amplitudes and
phases of the first ten harmonics of the sound are estimated and cancelled
 6 Automatic Classification of Pitched Musical Instrument Sounds 197

from the signal, in order to obtain a residual where the harmonics of the
second sound are analysed. The estimated spectra of the two sounds are then
fed to a neural classifier to recognize the two instruments.
Another approach to analysing duet signals is based on a so-called missing
feature theory that was developed for speech processing and speaker identifica-
tion. The main idea consists of using only the spectro-temporal regions which
are dominated by the target sound, and ignoring those that are dominated by
background noise or interfering tones. This approach is motivated by a model
of auditory perception, proposed by Cooke et al. [99], which postulates a sim-
ilar process in listeners. In polyphonic music, partials of one instrument often
overlap with those of another one. Consequently, the observed amplitudes of
these partials no longer correspond to those of any individual instrument.
Within the missing feature approach, these corrupted or unreliable features
can be excluded from the recognition process. The remaining information is
therefore incomplete, but the hope is that it is still sufficient to enable robust
instrument classification (additionally, it is possible to partially reconstruct
the missing values by exploiting known correlations between the missing and
the rehable values). Eggink and Brown [168], [167] used sub-band energies as
features, although other features could be utilized, too.

6.6.3 Classification of Instruments in Complex Mixtures

If classifying instruments in duets is still in its infancy, we should say that

identifying them in more complex polyphonic mixtures is a kind of newborn
subject. Recognizing all the instruments that are played at a given time in
a complex musical mixture is a tough problem. The pioneering proposal by
Kashino and Murase [331], which is closely related to sound source separation
(see Chapters 9 and 10), required that a representative waveform of each note
of each instrument to be identified was stored beforehand in order to per-
form a template adaptation and matching process. Once the onsets and FOs
were estimated (in their original system this was done manually), the most
prominent instrument tone was determined by comparing the mixture with
phase-adjusted example waveforms. In an iterative processing cycle, the en-
ergy of the corresponding waveform was then subtracted to find the next
most prominent instrument tone in the residual. The system also successfully
exploited high-level musical knowledge related to note transitions and voice
leading.
A closely related system was presented by Kinoshita et al. [343]. It in-
cluded improvements over the spectral template matching method of Kashino
and Murase [330]. FO estimation was performed prior to the instrument clas-
sification process to determine where partials of concurrent sounds overlap.
Features used for instrument identification were heuristically evaluated ac-
cording to their reliability and diagnosticity with relation to the stored tem-
plates of each instrument class. The overlapping features (e.g. power in bands)
were adapted according to the known templates, and those reliable and very
198 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

characteristic features of an instrument were selected for performing the clas-

sification. Both the system of Kinoshita et al. and that of Kashino and Murase
were evaluated using combinations of three instruments, and it is not clear if
they could be used with more dense mixtures.
Approaches other than source separation are feasible by considering some
simplifications to the problem. One of them consists of focusing on a given in-
strument or instrument type only. This is the case, for instance, when we con-
centrate on percussion instruments, which include very idiosyncratic sounds
(i.e., 'noises') that stand apart from the remaining 'harmonic' sounds gener-
ated by pitched instruments (see Chapter 5). Also, detecting a singing voice
above the rest of the instrumentation is an example of this approach [630].
Heittola applied a similar detection-without-separation strategy to detect the
presence of certain instruments (bowed strings, electric guitar, piano, saxo-
phone, vocals) in polyphonic music, but found that this straightforward ap-
proach did not produce very reliable results [281].
Deciding if a musical piece involves certain instruments in a 'predominant'
role, even though they cannot be accurately located in time, is another way to
address the identification of instruments in polyphonic music. An instrument
can be considered as predominant if it is present for a significant part of
a piece and it is louder than the others most of the time. For instance, in
popular music the singer's voice is usually louder than most of the other
instruments. Following this path, the systems proposed by Eggink and Brown
[170], and by Zhang [688] assume that FO estimation provides a rehable cue for
identifying a predominant instrument. The most dominant FO is determined
by a frequency-domain harmonic pattern matching process (see Chapter 8)
and the spectrum of the predominant instrument is then estimated and used
to compute discriminative features. Finally, a classifier that has been trained
using isolated sounds and solo phrases makes the decision on the instrument
class. It is important to note that in both systems it is required to decide on
only one instrument at each given time. In Zhang's system, an 'instrument
profile' was extracted for the whole piece by computing the proportion of notes
that were played by each instrument (hence, onset detection was performed).
In the case of Eggink and Brown, their system only attempted to identify
which instrument was, among 5 classes, the main one played in accompanied
sonatas and concertos (but not where it was played).
Identifying combinations of instruments directly (for example piano -h vio-
lin)^ instead of separating them, has been addressed by Essid et al. [180]. The
authors presented a multi-instrument recognition scheme capable of process-
ing jazz piano quartets based on prior knowledge of the musical context. Here,
the number of possible combinations can be reduced by building super-classes
consisting of unions of classes having similar acoustic features. The classifica-
tion can be performed hierarchically in the sense that a given test segment can
be first classified among the top-level classes, and then more precisely (when
needed) at a lower level. For example, if a test segment involves double bass,
drums, and trumpet, then it is first identified as belonging to the combined
 6 Automatic Classification of Pitched Musical Instrument Sounds 199

category 'Bass-Drums-MonoPitchedWind' and subsequently as 'Bass-Drums-

Trumpet'. A substantial amount of data from each of the effective sources
(i.e., individual instruments but also their usual combinations) is required to
train the classifier.^

6.7 Conclusions
In this chapter we have provided a review of the theoretical, methodological,
and practical issues involved in the automatic classification of pitched musical
instruments. Assigning instrument labels to analysis frames, sounds, or musi-
cal segments requires, first, a solid knowledge of the acoustic features of the
instruments, and of the ways we can exploit signal processing techniques to
convert them into numerical features. Additionally, a systematic methodology
comprising the collection of data sets for training and testing, the selection
and transformation of features, and the comparison of the results obtained
using different approaches, defines the right path for obtaining a robust clas-
sification system.
In the five years from our first review of the field of automatic classification
of musical sounds [283] to the present moment, the number of studies exceeded
twice the number of those published in 1990s. Some general tendencies can be
noted in these recent works: use of larger and more varied databases, interest
for unpitched percussion sounds, improvements in the methodological aspects,
and an increasing concern for practical applications and for dealing with truly
musical fragments.
The performance of the systems dealing with a large amount of isolated
sounds and pitched instrument classes achieves correct decisions nearly 70%
of the time, whereas a simpler decision on instrument family rises a bit beyond
80% accuracy. This leaves some room for improvement that could be achieved,
among other options, by carefully looking at the discriminative acoustic and
perceptual features that each class of sounds may have and then devising
feature extractors that capture them properly.
On the other hand, systems dealing with the classification of instruments
in musical excerpts are the actual hot-spot of the field even though the achieve-
ments are still quite modest. For the classification of solo phrases, the achiev-
able performance can be a bit better than that for isolated sounds, but when
duets or more complex combinations are considered, the performance drops
substantially. In those cases, systems become more complex as they rely on
multiple-FO estimation and on incomplete or noisy estimation of spectral and
temporal information. The current approaches try to avoid 'hard' source sep-
aration and exploit contextual or musical knowledge. This makes the problem
more manageable, even though the provided solutions have still been limited
in terms of sound combinations or musical styles.
^See also Chapter 5, p. 137, where the idea of recognizing combinations of sounds
directly is discussed from the viewpoint of percussion transcription.
200 Perfecto Herrera-Boyer, Anssi Klapuri, and Manuel Davy

In this review we have identified several open issues that could provide in-
teresting returns when properly addressed: (1) The need of a reference test col-
lection, containing enough variability in instruments, recording conditions and
performers to be considered as an unbiased sample of the real population of
instrument sounds, and granting that any proposed system can be fairly com-
pared to other alternative proposals. Fortunately, RWC is currently a serious
candidate that should gain wider acceptance among research groups. (2) The
need to develop better features and instrument-specific features. (3) The need
to investigate possibilities to embed some general knowledge about the task
into the classification system, such as the usual frequency ranges of the instru-
ments, voice leading rules etc., but also very specific knowledge, for example
by crafting ensembles of specialized classifiers. (4) The need to evaluate the ro-
bustness of a system under reverberant, noisy, or other distortion conditions.
(5) Incorporating instruments outside the typical orchestral ones: singing voice
and some electrophones, for instance, the electric guitar, would deserve spe-
cific studies by themselves, given their broad timbral registers. (6) The need
to the develop systems dealing with realistic polyphonic music signals.
The automatic classification of pitched sounds of musical instrument has
progressed a lot in the past five years. Even though real time [199], [415] and
commercial systems for instrument sound classification have been devised, ^^
we expect to find more of them soon, in connection with applied problems
posed by personal digital music players. This means that now is time to ex-
ploit the knowledge we have gained working with isolated sounds, in order to
address the identification of instruments played in polyphonic music. This is,
without doubt, the challenge for the forthcoming years.

Acknowledgements

The writing of this chapter was partially supported by the EU project SIMAC
(Semantic Interaction with Music Audio Contents) EU-FP6-IST-507142. The
first author wishes to acknowledge the input, feedback, and help received
during the preparation of the manuscript from Eduard Ay Ion, Emilia Gomez,
Fabien Gouyon, Enric Guaus, Eulalia Montalvo, Bee Suan Ong, and Sebastian
Streich.

^^http://www.musclefish.com, http://www.soundfisher.com,
h t t p : / / c u i d a d o s p . i r c a m . f r , http://www.audioclas.net,
 Part III

Multiple Fundamental Frequency Analysis

Multiple Fundamental Frequency Estimation
Based on Generative Models

Manuel Davy

LAGIS/CNRS, BP 48, Cite Scientifique, 59651 Villeneuve d'Ascq Cedex, France

Manuel.DavyQec-lille.fr

Western tonal music is highly structured, both along the time axis and along
the frequency axis. The time structure is described in other chapters of this
book (see Chapter 4), and it may be exploited to build efficient beat trackers,
for example. The frequency structure is also quite strong in tonal music. It
has been shown since Helmholtz (and probably before) that an individual
note is composed of one fundamental and several overtone partials [451], [193].
Though acoustic waveforms may vary from one musical instrument to another,
and even from one performance to another with the same instrument, they
can be modelled accurately using a unique mathematical model, with different
parameters.
In addition to a mathematical model that describes the waveform gener-
ation, the frequency structure of music can be used to derive priors over the
model parameter values. Here, we understand frequency structure in terms
of fundamental and partials structure, pitch/FO structure, etc. For example,
assume the instrument playing is a piano; then the note frequencies cannot
be just any frequencies; they have to match the piano key frequencies. Also,
the piano overtone partial frequencies are slightly inharmonic (that is, they
are not integer multiples of the fundamental partial frequency), and their fre-
quencies are described by a specific model [451], [193] which can be used to
build parameter priors. More generally, the structure of tonal music may be
exploited to build a Bayesian models that is, a mathematical model embedded
into a probabilistic framework that leads to the simplest model that explains
a given waveform (see Chapter 2 for an introduction).
The Bayesian setting is quite natural for this problem as it enables
the use of many heuristics within a rigorous framework. Moreover, acoustic
waveform models generally have many parameters, which cannot be accu-
rately estimated without regularizing assumptions, such as parameter priors.
Bayesian models for multiple FO estimation have received, however, relatively
little attention. A possible cause is that such models are complex, and this
makes their use difficultthough achievablewhen confronted with real data.
Sometimes they are also computationally heavy. However, such models enable
204 Manuel Davy

much more than multiple FO tracking. They do model the acoustic waveform:
the parameters which are estimated from real musical records may be used
for multiple FO estimation, but also for monaural source separation, sound
compression, pitch correction, etc.
In this chapter, we present several approaches to multiple FO estimation
that rely on a generative model of the acoustic waveform. More precisely,
we present a noisy sum-of-sines models which has been studied by many
authors in various contexts. This model and some of its variants are pre-
sented in Section 7.1. Section 7.2 introduces a Bayesian off-line processing
method which requires notewise processing (that is, processing is performed
on a complete waveform section which does not include note changes). In Sec-
tion 7.3, we present the on-line processing model of Cemgil et al. [78], Dubois
and Davy [158], and Vincent and Plumbley [646]. Section 7.4 is devoted to
other on-hne multiple FO tracking algorithms that rely on incomplete or in-
direct acoustic waveform modelling: for example, the approach of Thornburg
et al. [625] and Sterian et al. [609] models the time evolution of time-frequency
energy peaks. Dubois and Davy [159] model the signal spectrogram (which
comes down to estimating on-line the acoustic waveform up to the initial phase
parameter, though). Section 7.5 presents some conclusions.

7.1 Noisy Sum-of-Sines Models

In this section, we first present some simple models which were developed for
single FO acoustic signals. The earliest noisy sum-of-sines acoustic waveform
models were developed for speech synthesis; see e.g. [449]. These models did
not assume, however, frequency relations between the fundamental partial and
overtone partials. Laroche et al. introduce a harmonic plus noise model [394]
which assumes such relations. These models were soon used for music process-
ing; see [575] for a review of early methods.

7.1.1 Single FO Stationary Models

The frequency structure of tonal music acoustic waveforms has been observed
for many years. As can be seen in Fig. 7.1, these waveforms are almost periodic
and their Fourier transforms reduce to (approximately) sums of sine waves
whose frequencies are multiples of a given frequency. For a perfectly periodic
(infinite length) signal x, with discrete time n = 1, 2 , . . . ,
M
^(^) ~ / ^ o^sin(27rmA:in) -f a^cos(27rmfein). (7.1)

In the following, we name the sine + cosine component with frequency ki

the fundamental and components with harmonic number m = 2 , . . . , M the
 7 Multiple FO Estimation Based on Generative Models 205

overtone partials} The model in (7.1) is quite simple, but it is rather theo-
retical. Real signals always include components which cannot be modelled as
individual sines or cosines: for example, a flute player breathing can be heard
in recorded signals, and this is highly non-periodic [219]. As such components
are quite diflPerent from one occurence to another, they can be jointly modelled
in terms of their statistical distribution as a noise component e, yielding the
model
M
x[n) 2_\ ^^ sin(27rmA:in) -[- a^ cos{27rmkin) + e(n). (7.2)
771=1

100 150
Time (ms)
Fig. 7.1. Flute acoustic waveform (top) together with its spectrogram (bottom).

Two noise statistical models have received some attention. The simplest
assumes e(n) to be a white noise with Gaussian distribution (see for exam-
ple [394], [616], [124]). This is the less informative assumption, as in this case.

^As pointed out in Chapter 1, the frequency of the fundamental partial, denoted
here by /ci, is different from the fundamental frequency FO, which is the inverse of
the acoustic waveform period.
206 Manuel Davy

e(n) is a purely random sequence with a flat power spectrum. Another popu-
lar model takes the form of an autoregressive process [576], [575], [292], [609],
[122] (see Chapter 2 for a presentation of autoregressive models), which also
corresponds to random sequences, but with non-fiat power spectrum. A review
of possible choices for e may be found in [310], [308].
In addition to non-harmonic components, acoustic waveforms produced
by real musical instruments also have another important characteristic: they
are not strictly periodic. This is explained by two phenomena: inharmonicity
(or partial de-tuning) and partials amplitude nonstationarity. Inharmonicity
appears whenever the frequency of the partial with harmonic number m is not
exactly mki. For the example in Fig. 7.2, where several periods of the acoustic
waveform produced by a piano, a flute, and a clarinet are superimposed, the
non-periodicity appears clearly. This may be caused by amplitudes decay (for
all three examples), but also by inharmonicity (piano example). Note that in
Fig. 7.2, two of the three periods plotted are contiguous, whereas one is taken
further apart. A more general model enabling inharmonicity is
M
x{n) = 2_2 ^^ sin(27rA:mn) -f a^ cos(27rA:mn) + e(n), (7.3)
m=l

where the partial frequencies km (m = 2 , . . . , M) are related to ki by a more

elaborate relation than km = rnki. For example, Fletcher and Rossing [193,
p. 363] propose the following piano inharmonicity model:

Kri mki (7.4)

1+B '
with B e [10-^10-2]. For B = 0.0004, this shifts the 17th partial at the
frequency position of the 18th partial.

3.8 0 20 7
Piano (FO 262 Hz) Flute (FO 490 Hz) Clarinet (FO 135 Hz)
Fig. 7.2. Three superimposed periods of the acoustic waveforms played by a piano,
a flute, and a clarinet. The waveforms are not strictly periodic and the three periods
represented (in solid, dashed, and dashed-dotted lines) are not exactly superimposed.
The time scale is in milliseconds.
 7 Multiple FO Estimation Based on Generative Models 207

Aside from partial frequency models, it may be useful to model the partial-
to-partial amplitude profile (referred to as the spectrum envelope in the follow-
ing) . The power spectrum of a note is formed by the instrument body response
which modulates the partial frequency peaks. Of course, this modulation de-
pends on the instrument and it may be characterized by a spectrum envelope
(see the saxophone family example [193, p. 497] ) which may be modelled.
Spectrum envelope models need to be quite flexible, though, because some
instruments have special behavior. For example, clarinets have almost zero
amplitude for every other low frequency partials (partials with even harmonic
numbers m); see Fig. 6.2 p. 166. Godsill and Davy [213] propose a statisti-
cal model where the amplitudes are assumed to be approximately constant
below some cut-off frequency /Ccutoff, and decay exponentially for higher fre-
quencies. The parameters defining the model (fccutoff and the exponential decay
rate) are to be estimated from the processed acoustic waveform. Cemgil [78]
uses an exponential decay where the amplitude a^ of partial m equals [a|]
(and similarly with a ^ ) . Alternative models may be proposed, based on the
spectral smoothness principle; see Klapuri [351].

7.1.2 Single FO Non-Stationary Models

The models presented above permit quite good modelling of very short signal
portions, insofar as the amplitudes and frequencies do not vary too much over
time. As explained in Chapter 4, however, the amplitudes do vary quickly
enough so that the above models cannot be used to process musical segments
longer than 20 to 50 ms. A more general non-stationary model is
M
x{n) = ^ c^m(^) sin (27r/Cm(n)n) + a ^ ( n ) cos {2TTkm{n)n) + e(n), (7.5)
771=1

where the amplitudes c^%j^{n) and a^{n), the frequencies A:m(^) for m =
1 , . . . , M , and the noise statistics now depend on the time. This model is
quite flexible, but it is no more a sine-plus-noise model: to understand this,
assume km{n) is an independent sequence of random frequencies with M = 1;
then the wave generated is not a sine wave at all! This shows that the model
in (7.5) should be constrained so as to be suited to tonal music. This can be
done in many ways, but the simplest is certainly to assume a smooth time evo-
lution of the amplitudes and frequencies, and assume a statistically (almost)
stationary noise.
Many amplitude and frequency evolution models may be found. The am-
plitude evolution model should mimic the way notes appear and disappear
in music (onset and decay), whereas the frequency models should adapt to
stationary cases (for instruments such as the piano where the performer has
quite limited influence on the note frequencies evolution) or vibrato (e.g., for
violins). Many such models may be found in the music synthesis literature,
and they are generally quite instrument specific; see e.g. [543], [628]. These
208 Manuel Davy

models may be used for musical signal analysis, in particular when the analy-
sis conditions are well controlled. Here, we present more general models which
can adapt to different instruments and different kinds of music.

Amplitude Evolution Models

As pointed out above, amplitude evolution models need to allow quick enough
variations in order to fit note onset and decay. However, amplitudes should not
vary too quickly: to make this point clearer, assume the acoustic waveform is
a sine wave and let M = 1 and fci = 0. Then, the only way for the model to fit
the data is for the amplitude itself to be a sine wave. This illustrates that, given
an acoustic waveform, there is not a unique frequency/amplitude parameter
set for the model in (7.5); rather, there are many. It is thus important to
prevent the time-varying amplitudes from fitting the sine waves. This can
be easily done by selecting amplitude evolution models that do not permit
oscillations with frequencies over some 10 Hz, for example.
A relevant model is that of damped amplitudes^ where the amplitude
evolves according to a decreasing exponential; that is (where we drop the
partial index m for notation clarity),

a^(n) = a^exp(-An) and a^{n) = 5^exp(-An), (7.6)

where the damping factor A tunes the amplitude decay rate, and a^, 5^ are
fixed initial amplitudes. When substituted into (7.5), this yields a damped
sinusoids model, as used by Hilands and Thomopoulos [292] for multiple si-
nusoids frequency estimation or by Cemgil et al. [78] for music transcription;
see Section 7.3 below. Note that such an amplitude time evolution may be
coupled with a spectrum envelope as described in Section 7.1.1.
Another amplitude evolution model consists of assuming a random walk
or an autoregressive process, as proposed in [106] for chirp signals. This re-
duces the amplitude parameters to the set of AR coefficients, which may be
small. These coefficients should be chosen, however, so as to ensure a smooth
amplitude.'^
The last model presented here writes the time-varying amplitude as a sum
of weighted smooth, time-localized functions with time-domain shape (j){n)
I I

a%n) = ^ a ^ 0 [ n - z A n ] and a'^in) = ^ 5 J ( / ) [ n - iAn], (7.7)

7=0 7=0

where af, a^ are the amplitudes (also called weights) associated to each time-
localized function (f)[n iAn]- The step A^ sets the spacing between two
such successive functions. The shape (f) is typically chosen so as to obtain
smooth amplitude profile; in general it is one of the standard 'shding windows':
^This can be obtained by choosing the AR coefficients whose corresponding char-
acteristic polynomial has zeros with relatively small amplitude [339].
 7 Multiple FO Estimation Based on Generative Models 209

Gaussian, Hamming, Hanning, etc. Here also, the amplitude evolution from
one frame to another may follow a random walk [159] or may be unrelated
a priori [616]. An important remark is that the resulting time-varying sum-
of-sines model is closely related to a Gabor representation^ where a signal is
decomposed into windowed sine/cosine waves whose time-frequency locations
are determined by a regular lattice [184]. Here, the lattice is regular along
the time axis and irregular along the frequency axis. Gabor-style amplitude
models were used by Davy and Godsill [122] (irregular lattice) and Wolfe
et al. [676] (regular lattice).

Frequency Evolution Models

A simple frequency evolution model may assume constant frequencies. This

is actually quite a realistic assumption insofar as that the instrument player
has no means of changing note frequencies during their emission. This is the
case for many instruments, including pianos, flutes, oboes, etc. It can also be
a sought-after effect to keep it constant in other instruments such as violins,
cellos, or trumpets.
Aside this simple model, a possibility is to assume random walk frequency
evolution, that is

km(ji) = km{n - 1) + Vm{n) forn = 2 , . . . , r , (7.8)

where Vm{n) is a Gaussian white random noise with some fixed variance.
In order to be valid, though, the noises Vm{n), m = 1 , . . . , M have to be
correlated so that the frequencies of related partials follow similar evolutions.
In the case of abrupt frequency changes, the models presented here are
out of their validity domain. However, a standard assumption is that no note
changes occur within the acoustic waveform segments processed.

7.1.3 Multiple FO Non-Stationary Models

It is straightforward to extend from single to multiple FO models by simply

adding J single FO models. Equation (7.5) becomes
J M,
^(P) = X I X I ^i,m(^) sin {2'Kkj^rn{n)n) -j- Q^j^^{n) cos (27r/c^-,^(n)n) + e(n),

(7.9)
where the amplitudes and frequencies are modelled as in the single FO mod-
els discussed above. In specific contexts, it may also be useful to model links
between the frequencies and amplitudes of different notes. For example, the
sound of an electromechanical organ may be modulated by rotating loud-
speakers; the strings of an electric guitar may be jointly tightened by a moving
bridge. Such links may be either deterministic or probabilistic.
210 Manuel Davy

The mathematical models described above are characterized by many pa-

rameters to be estimated. Depending on the specific model selected, the num-
ber of parameters for a 100 ms excerpt with 3 notes may vary from about 50
to more than 3000. In order to make the estimation feasible and non-trivial,
it is necessary to define a measure of the fit between the data and the model
and to favour simple models. This can be easily done with a Bayesian model.
The following two sections define such probabilistic models first in the off'-line
case, and then in the on-line case.

7.2 Off-line Approaches

The off'-line approaches are characterized by their global viewpoint on the

parameter estimation problem. In the Bayesian estimation settings, we need
a likelihood term (which describes the model-to-dat a fit) and priors, because
the quantity of interest (the posterior) is proportional to the product of the
former two terms; see Chapter 2. An important assumption is made in this
section: the processed acoustic waveforms have been segmented so that they
do not include transitions between the notes. This can be achieved by applying
some onset detection technique; see Chapter 4.

7.2.1 Likelihood

The likelihood term is completely described by the mathematical model se-

lected, and by the probability density function of the noise e(n). As an illus-
trative example, we assume here that this noise is zero-mean white Gaussian
with variance cr^, and we assume the model of (7.9) with the amplitude
evolution model of (7.7) and constant frequencies. In order to simplify the
notation, we write the waveform mathematical model in vector form. Let
X = [x(l) . . . x{T)]^ and e = [e(l) . . . e{T)]^. Moreover, we define a to be
the vector of sine/cosine amplitudes. Finally, let D be a matrix with T lines,
whose rows contain the sine/cosine Gabor atoms for each note partial fre-
quency, in the same order as in a (see [122], [124] and (2.68) in Chapter 2 for
explicit definitions of a and D). With these notations, the acoustic waveform
model can be written as
X = D a + 6. (7.10)
All of the models presented above may be written in this vector form,
where the basis function matrix D and the amplitude vector a may have
different structures. In any case, the dimension of D depends on J, on M =
[ M l , . . . , Mj], and on / , and its rows depend on the frequencies k (where k is
the vector that contains all the frequency parameters kj^rni'^)^ m = 1, ., Mj
and j = 1 , . . . , J ) .
It is now possible to write the likehhood of this model. Of course, it depends
on the probabilistic assumptions made about the noise e. Let us define x and
 7 Multiple FO Estimation Based on Generative Models 211

D such that x = x and D = D whenever is a zero-mean white Gaussian

noise of variance a^ and x == Ax, D = A D if the noise is autoregressive of
order p and coefficients a = [ a i , . . . , ap], with Gaussian noise of variance a^.
The matrix A is defined as
1 0 .. 0
a\ 1 0 0

A = -flp ... -ax 1 0 . .. 0 (7.11)

0

: '. . 0

0 ... 0 a^ ... ai 1
Using these notations, the likehhood is written

p(x|J,M,a,k,CT^(a)) = {2Tra^)~'^/^ exp Ix - Dal (7.12)

2a2
where the possible AR coefficients a are only included in the likelihood when
e is autoregressive. In the following, the optional parameter a is omitted for
the sake of simplicity.

7.2.2 Prior Distributions Selection

The next step towards a Bayesian model is that of prior distributions selection.
Priors are essential in this problem, because they make the estimation of the
numerous parameters feasible: without priors, many possible solutions would
be likely, including trivial, meaningless, and over-complex ones. In particular,
priors are useful to avoid having the model feature too many notes. There are
many possible priors for J, M, a , k, cr^ (and possibly a). However, compu-
tational load is reduced when selecting a Gaussian amplitude prior p(a) and
an inverse gamma prior distribution for the variance parameter cr^. Thus, the
overall prior should be written as
p ( J , M , a , k , a ^ 1 0 ) = p(a|J,M,k,a2|V^)p(J,M,k|V^)p(a2|V^), (7.13)
where t/? is called the hyperparameter vector-^ it contains the prior parameters,
called hyperparameters ([122], [124] and Chapter 2). Similar to (2.74) and
(2.75) in Chapter 2, we select here
p ( a | J , M , k , a 2 | ' 0 ) = A/'(a;0,(j^Z'c), (7.14)

P(^'IV^) (7.15)

which enables us to write the posterior as a product of three terms

p(J,M, a,k,<T2|x,'0) oc (7.16)
P(CT2| J, M , k,X, 0) p ( a | J, M , k, CT^ X, 0) p(J, M , k|x, tp).
212 Manuel Davy

The three terms of the posterior distribution are

p ( a 2 | J , M , k , x ) = IQ ( . 2 ; ^ , f^o|_fH) , (7.17)

p(a|J,M,k,(T2,x) =Ar(a;SD"rX,(72S), (7.18)

T + i^i

p ( J , M , k | x , V;) a i A _ _ _ L _ _ p ( J , M , k | i / > ) , (7.19)

where the matrices S and P are defined as in Chapter 2, page 47.

Equations (7.17), (7.18), and (7.19) define the posterior, up to the prior
p(J, M,k|'0) which still needs to be selected. The parameters v^^ v\ tune the
prior over the noise variance a^ and they should be set so as to favour small
values of G^ . This ensures that the residual x - D a is as small as possible.
This can be achieved by selecting i/g < < 1 and v\ \ (typically, smaller
than 10-2).
The covariance matrix J^ot controls the amplitude coefficients. Standard
choices are
The g-prior^ which consists of setting Ua = 7^(6^13) . This prior has
two advantages: First, the amplitudes of sinusoids with very close frequen-
cies are constrained to small values. This shares the energy between those
two neighbour components. Second, 201ogiQ(7) can be interpreted as the
expected signal-to-noise ratio (in dB). Third, the numerical computations
are even easier because the matrix S becomes really simple. The hyperpa-
rameter 7 should be considered unknown, and might be estimated. Actu-
ally, its value may be quite critical for correct estimation of the number of
notes/partials. A standard choice consists of selecting a prior p(72) as an
inverse gamma distribution [21], [656], [122].
A block diagonal matrix, made of 2 x 2 blocks. This choice enables the
implementation of a fast estimation algorithm [127], [124]. Each block
I7cK[n] located at lines/rows (2n,2n + 1) can be set as, e.g., i7a[n] =
7^(D^Dn) , where D^ denotes the two rows of D located at indexes
2n and 2n + 1. The hyperparameter 7^ may be set to the same value for
all blocks, which assumes the same prior amplitude for all partials and all
notes. It may also have the same value for all blocks that are related to
the same note and have different values from one note to another. Finally,
7^ may also be used to model the amplitude decay from one partial to
another, according to the given spectrum envelope; see Section 7.1.1.
A diagonal matrix, which may be proportional to the identity, i.e., Ua =
7^1, where 7^ is an unknown hyperparameter with inverse gamma prior,
or Sot is a diagonal matrix where each diagonal term 7^ is set for each
partial, according to some spectrum envelope.
 7 Multiple FO Estimation Bcised on Generative Models 213

When additional prior information is available (for example, the instru-

ments playing are known), it is possible to design an even more specific prior
over the amplitudes by setting the 7^'s to some preset values, multiplied by
a common scale parameter to be estimated.
The remaining parameters joint prior p( J, M, kit/?) should also be selected.
This prior may be decomposed into individual terms as

p(J,M,k|i/.) = p ( k | J , M , V ) p ( J , M | V ) . (7.20)

Similar to the amplitude prior, the frequencies prior may have various shapes,
depending on the level of prior information available. A very general prior is

M,
p(k|J,M,V^) = l[ p{kj^i\Mj,ip) Y[ Pikj^mlkj^i.'ip) (7.21)
j=i m=l

In (7.21), the fundamental partial frequency prior p{kj^i\Mj,ip) may be in-

strument specific. For instruments with keys (pianos, clarinets, etc.) this prior
may have the shape depicted in Fig. 7.3. In the general case, this kind of prior
may also be used to perform frequency quantization (see Chapter 12); in this
case, considering the A4 frequency as an additional unknown parameter to
be estimated around 440 Hz makes the quantization more robust. Another

mHMAAAAAAAAAA
Al BlCl Dl ElFl Gl A2 B2 C2 D2
Frequency (note)
Fig. 7.3. Prior distribution of the fundamental frequencies p(/cj,i|Mj,i/?) in a 'key
instruments' model. The spread of the Gaussian function increases with frequency,
i.e., with the note label in {A1,A1#,B1, ... }. This is aimed at limiting areas where
the prior is zero, in order to make the numerical estimation easier.

efficient fundamental partial frequency prior is the uniform distribution over

the interval [0, ks/2Mj] {j = 1 , . . . , J ) . The upper limit ks/2Mj ensures that
no overtone partial has a frequency higher than the Nyquist frequency ks/2,
as this would cause aliasing.
The overtone partial frequency prior terms p(A:j,m|^j,i,'0) need to take
possible inharmonicity into account.
A good, instrument unspecific, partial frequency prior is given by a trun-
cated Gaussian distribution

P(^j,m|A:i,i''0) = ^{^j,m;mkj^i,{mkj^ifal^)lrk^ (7.22)

214 Manuel Davy

where cr^^ is a small variance parameter (which can be chosen equal for all
partials and all notes). In (7.22), the indicator function I (see p. 28) restricts
kj^rn to the range [-^, - ^ ] to avoid overtone partial frequencies switching.
The last prior term to be defined, p(J, Mlt/?), is also the most critical.
This term should be strong enough to avoid models with too many notes and
too many partials. However, it should not overly penalize models with many
notes and partials, because some important partial/note components may be
missed.
Overall, there are two standard choices for p{J,M\ip) (others may be de-
signed, though). The first consists of using the hierarchical structure p( J, M | ^ )
= p(J|t/?) n / = i Pi^jW^ where each term p{Mj\xp) is a Poisson distribution
with parameter Aj

p{Mj\iP) = P ( M , ; A , ) = e - ^ ^ ^ , (7.23)

and with p(J|'0) = V[J] A). The hyperparameters Kj {j = 1 , . . . , J ) and A

play a crucial role in the estimation. Indeed, a model with too many allowed
partials may tend to underestimate the fundamental partial frequency by one
or several octaves, that is, by setting it to, e.g., half its true value. This
adds fake partials in between the existing ones. It is thus important to let
the hyperparameters A j (j = 1 , . . . , J) and A adapt to the analysed acoustic
waveform. A robust approach consists of defining gamma priors for these
hyperparameters; see [124].
The second approach consists of defining a prior for the model total number
of partials R = Y,j=i ^j- Here again, a Poisson prior p( J, M|t/?) = p(i?|'0) =
7^(i?;A) may be used, where the hyperparameter A is considered unknown
with a gamma prior.
The probabilistic model being completely defined, the important issue is
now to design an efficient estimation algorithm.

7.2.3 Estimation Issues

Parameter estimation from a Bayesian model may be achieved in many ways.

The main two approaches are minimum mean squared error (MMSE) and
maximum a posteriori (MAP) estimators (see Chapter 2, page 41). Here,
the structure of the posterior distribution p(J, M , a , k , cr^|x, i/?) is especially
complex, with discrete parameters ( J and M), positive parameters (k and
(j^), and real-valued vector parameters ( a and the optional parameter a).
A possible solution consists of estimating J and M by marginal MAP, that
is, by selecting the number of notes JMAP and number of partials MMAP for
which p(J, M|x,-0) is maximumthis is typically computed by Monte Carlo
techniques; see Chapter 2. Then, the other parameters may be estimated by
marginal MMSE, conditional to JMAP and MMAP, that is,
 7 Multiple FO Estimation Based on Generative Models 215

[a, k, a%us^ = ^,^^,i,,,2^^j^,,,M^^^,^) ["' k, ^ ' ] (7.24)

This kind of estimator requires, however, the computation of marginal distri-

bution, of maxima, and of integrals such as the expectation in (7.24). This
requires dedicated algorithms, as outlined in the next section.

7.2.4 Computational Issues

The computation of estimates from the posterior p(J, M , a , k , cr^|x, V^) re-
quires the 'exploration' of this multidimensional probability distribution. Sev-
eral works address similar problems. Andrieu and Doucet [21] propose a
Markov chain Monte Carlo (MCMC) algorithm for a simple noisy sum-of-
sines Bayesian model. Walmsley et al. [657], [658], [656], derive an MCMC
algorithm for single FO Bayesian models. Finally, Davy and Godsill [122],
[126], [124] propose various MCMC approaches for multiple FO models, with
a fast implementation [127], [124].
As explained in Chapter 2, the aim of MCMC algorithms is to produce
a chain of samples J^'\M^'\oL^'\k^'\a'^^'\^'<'\ for i = 1,2,3,.... These
samples are used to estimate the various quantities described in Section 7.2.3
above. The derivation of such algorithms being quite lengthy, the interested
reader may refer to the publications cited for details. We present below (Al-
gorithm 7.1) the general structure of an MCMC algorithm dedicated to a
multiple FO noisy sum-of-sines model, with a two-level hierarchical structure,
from the individual partial level up to the multiple-note level.

Algorithm 7.1: Overall MCMC Algorithm for Multiple FO Models.

Initialization.
Step 7 . 1 . 1 Initialize the parameters J, M , a , k, a^, and ip.
- Sample tjA^^ from its prior distribution.
~ Sample J^^^ according to some initial distribution qinit(-^)-
- For j = 1 , . . . , J^^^ sample Mj according to its Poisson prior distribution.
~ Sample k^^^ according to qinit(k|x) where qinit(k|x) is the probability distribution
proportional to the Fourier spectrum of x (see [21] for a similar implementa-
tion).
-- Sample thejioise variance parameter a^^^^ according to its posterior distribution
p(a2|/^),M('\k(^\(a(^)),x,'0(^)) given in (7.17).
-- Sampje tjie amplitudes a^^^ according to their posterior distribution
p ( a | j ( ^ \ M ( ^ ) , k ( ^ \ a 2 ( i ) ^ ( ~ ( i ) ) ^ ^ ^ ^ ( i ) ) gj^gn j ^ (7;^g)_
F o r i = 1,2, . . . , i V , do ^ _
Step 7 . 1 . 2 Sample the note parameters J^'\ M^'\ k^'^ .
- With probability / x j , try to add a new note using a note birth move, which
consists of generating a set of note parameters (number of partials, frequencies,
amplitudes), and testing it using a Metropolis-Hastings reversible jump.
216 Manuel Davy

- Otherwise, with probability i/j, try to remove a note using a note death move,
which consists of selecting one of the existing notes at iteration i 1 and testing
its possible removal using a Metropolis-Hastings reversible jump.
- Otherwise, with probability 1 fij uj, try the note update move as follows:
Set/*) ^j^'~y.
For j = 1 , . . . , J^*), update the parameters of note # j by possibly changing
the number of partials or their frequencies and amplitudes, yielding M^,
k] ^ and a} \
Step 7.1.3 Sample g^ from p(a^| J^*\ M(^^k^*^x); see (7.17).
Step 7 . 1 . 4 Sample -^^'^ from its posterior distribution (either directly, or us-
ing a Metropolis-Hastings test).
- Set 2 ^ z + 1.

Algorithm 7.1 deserves several comments. First, the hierarchical levels ap-
pear clearly. The note death/birth/update moves correspond to the highest
level (that of notes). The middle level corresponds to adding/removing partials
inside a given note, or changing its fundamental partial frequency. The lowest
level is that of individual partials, whose frequencies and amplitudes may be
updated. Second, the Metropolis-Hastings moves as well as the initialization
require proposal distributions q(-)- In order for the distribution of the samples
j(^)^ M^^), a^*\ k^^\ 52(^)^ -0() to converge quickly to the posterior distribu-
tion, these proposals may be built on heuristics. For example, the spectrum of
X is used in order to build the frequencies proposal distribution. Other similar
heuristics may be included in the same way; see [124]. Overall, such MCMC
algorithms may be seen as a rigorous way to use various heuristics for multiple
FO estimation.

7.2.5 Performance

The above class of Bayesian models being based on a generative model, they
can be used for many tasks, including multiple FO estimation and signal com-
pression. The performance estimation thus depends on the task assigned to
the algorithm. The computation for the full inference problem reported in
Davy et al. [124] is about 1.35 seconds per MCMC iteration for 0.5-second
excerpts, where 800 iterations are necessary on average.
In terms of FO estimation, Davy et al. [124] report 100% accuracy for
single FO (that is, for J = 1), about 85% for J = 2, 75% accuracy for J = 3,
and 71% accuracy for J = 4. The test samples were random mixes of single
FO acoustic waveforms from Western classical music instruments. In terms
of residual energy and reconstruction accuracy, all experiments showed very
good reconstruction,^ in spite of some octave errors.
"^It is much more difficult to evaluate the reconstruction accuracy with a psychoa-
coustically relevant measure. The residual total energy may yield such a performance
estimator, though.
 7 Multiple FO Estimation Based on Generative Models 217

7.2.6 Extensions

The above models may be extended in many ways. For example, it is possible
to use binaural (stereophonic) records in order to make the estimation more
robust [120]. This is implemented by defining two models, one for the left
channel and one for the right channel. These models are connected through
their parameters prior. This can be extended to multi-track recordings, as in
standard source separation approaches; see Chapter 9.
Another possible extension comes from a sequential view of music process-
ing. The off-line approach presented here concerns music waveform segments
which come from a longer excerpt, though they could be applied framewise.
It is possible to use the parameter estimation results of previous segments to
design the prior for the current waveform segment, in terms of instruments
playing, A4 frequency tuning, etc.

7.3 On-Line Approaches

The off-line approaches described above have the major advantage of being
quite accurate, because they have the full acoustic waveform available for
processing. The drawback is that the computational requirements may be
high. On-line approaches, on the other hand, only use the current acoustic
waveform sample x{n)or the current frameat each processing step n.
They also use information from the past, thanks to sequential priors (also
termed transition probability distributions) over the model parameters. Fi-
nally, it should be noted that sequential models are quite attractive for music
transcription, because they do not require a separate onset/offset detection
mechanism, and they provide some flexibility to formulate completely on-line
inference schemas, such as those presented in this section.
In this section, we first introduce the approach of Cemgil et al. [78],
and then we describe two approaches based on an explicit sliding window-
based model, proposed by Dubois and Davy [158], [160] and Vincent and
Plumbley [646]. Note that Irizarry [308] also proposes such a model, but it is
restricted to single FO.

7.3.1 Approach of Cemgil et al.

The noisy sum-of-sines models presented above are written in a non-sequential

form. Sequential approaches, however, require the noisy sum-of-sines model
to be written in a sequential form.

Sequential Noisy Sum-of-Sines Model

Cemgil et al. [78] propose to write a sine/cosine wave with frequency k and
unity amplitude as follows:
218 Manuel Davy

e{n) = B(fc)6>(n - 1), n=l,...,T, (7.25)

where 6{n) is a two-dimensional vector and B(/c) is a rotation matrix written

as
B(^) = cos(27rfc) sin(27rA:) (7.26)
sin(27rA:) cos(27rfc)
which rotates the vector 0{n 1) by 27rk radians counter-clockwise. The first
dimension of 6 carries the cosine component, whereas the second carries the
sine component.
It is now possible to write many sum-of-sines models in a sequential form,
including time-varying amplitude/frequency cases. Cemgil et al. [78] focus on
a damped amplitude, constant frequency model as

0(n) = X{n)B{k)e{n-l), (7.27)

x{n) = [1 0] 0(n), (7.28)

where A is the damping factor that tunes the amplitude decay rate, with
0 < A(n) < 1. The vector [1 0] is used to project the vector 0{n) onto one
axis; see Fig. 7.4. The initial phase is tuned by the initial vector 0(0) at time
0. In order to model the note partials, and their frequency/amplitude rela-

Rotating vector 0(n) Projection: damped sine x{n)

Fig. 7.4. The model based on a rotation matrix B(/c) rotates the vector 0{n) by
27r/t radians counter-clockwise, and applies a damping factor. When projected onto
one axis (here, the sine axis), this produces a damped sine wave with frequency k
(adapted from [78]).

tions, the vector 6 may be extended to M sines/cosines, that is, 9{n)^ =

[01 (n)^, 02(^)^ 0 M ( ^ ) ^ ] (of dimension 2M), and the sequential model be-
comes

0{n) = A{k)e{n-l), (7.29)

x{n) = C0(n), (7.30)
 7 Multiple FO Estimation Based on Generative Models 219

where C = [1 0 1 0 . . . 1 0]. The 2M x 2M matrix A(A:) is defined as

'Xi{n)B{ki) 0 0
0 A2(n)B(/c2) :
A(fc) (7.31)
: *. 0
0 .-. 0 XM{n)B{kM)_
Cemgil et al. assume harmonic frequency relationships of the partials and
exponentially decaying spectrum envelope, from one partial to another. In
other words, km '^k\ and A^(n) = A'^(n) for m = 1, . , M, where M is
assumed to be known.
Prom this sequential setting, Cemgil et al. derive a probabilistic model.
First, it is assumed that the note frequencies belong to a frequency grid with
K nodes. The note fundamental partial frequencies are assumed to be one of
the K predefined frequency positions among, for example, note frequencies of
the tempered scale over several octaves. Second, each of the grid frequencies is
assumed to be in one of the two states 'mute' and 'sound' at each time instant.
In the following, we denote by e/c(^) the state of the frequency grid point with
index k at time n, and we have either e/e(n) 'sound' or e/e(n) = 'mute'.
Coming back to the model in (7.29)-(7.31), the 'sound' case corresponds
to a damping parameter, denoted A^""""^, and the 'mute' case corresponds
to another damping factor A"""*^, with A^""""^ > A"'"*^ From these damping
factors, Cemgil et al. define two matrices A(/c,efc(n)) for efc(n) = 'sound' and
efc(n) = 'mute' by replacing A(n) in (7.31) with either A"^"*^ or A^""^ The
probabilistic model is written for each of the K note frequencies in the grid.
For n = 1 , . . . , T and for a frequency k in the grid,
efc(n) ~ p(efc(n)|efc(n-l)). (7.32)
Moreover, if no onset occurs, that is, if ek{n) = efc(n 1) or if efc(n) =
'mute' whereas efc(n 1) 'sound', then the damped sum-of-sines model
with damping factor A^""^ or A"""'^ is active:
0fc(n) - Ar((9fc(n);A(/c,efc(n-l))0fc(n-l),2;__t), (7.33)
or, if an onset occurs, that is, if efc(n) = 'sound' whereas efc(n 1) = 'mute',
then the amplitude of the sine wave at this frequency is re-initialized by sam-
pling a new initial vector 0fc(n) as follows:
e^{n) ~ Ar((9fc(n);02M,^onset). (7.34)
The acoustic waveform is modelled overall as the sum over all the frequencies
in the grid as
x{n) = ^ C 0 f c ( n ) + e(n), (7.35)
k

where e(n) is a Gaussian white noise with variance cr^. Equation (7.35) yields
the model likelihood at each time n (also termed the observation probability
220 Manuel Davy

distribution)^ and the sequential priors given in (7.32)-(7.33) enable writing

of the posterior in a sequential form. The multiple FO model being completely
defined, we next describe the estimation algorithm.

Parameter Estimation and Algorithmic Issues

Cemgil et al. propose to implement MAP estimation, that is, estimate the
sequence e^ = [e;c(l),..., e/c(T)] for each frequency k in the grid, where e^^ is
called a piano roll, as explained in Chapter 1. Note that the piano roll contains
all the necessary information for multiple FO estimation. Let us denote by
e^MAP the estimated piano-roll sequence. It is computed as

efCMAP = argmaxp(eK,'0|x), (7.36)

where the hyperparameter vector ip includes cr^, 27onset, ^no onset, A^"'"''^ and
^mute rYYie postcrior p(eK|x) is obtained by integrating out the parameters
6k{n) for all k and all n = 1, ., Tdenoted by the shorthand 6K{^ T)as
follows:
p(eK,i/^|x) = / p ( e K , 0 K ( l : T ) , V ^ | x ) d ^ x ( l : r ) , (7.37)
Je
where the full parameter posterior is given by Bayes's rule

p(eK,0K(l:T),t/;|x) ex p(x|eK,^K(l : T ) ) p ( e K , 0 x ( l : T)|V')p(^).

(7.38)
The terms p(x|ex,^/^(l : T)) and p ( e x , ^ K ( l T)\xp) are computed sequen-
tially as follows:
T
p ( x | e K , ^ K ( l : T ) ) = l[X{x{n);J2C0k{n).cT^), (7.39)
n=l k

T
p{eK.OKil:T)\xP)=l[ p(efc(l), 0^(1)) n P(efc(^)|efc(n - 1), t/?)
k
T 1
X
n=2
(7.40)
In (7.40) above, the transition pdf p[Ok{n)\0kin 1),'0) niay correspond to
either the 'onset' given in (7.34) or the 'no onset' pdf, (7.33). The initial pdf
of the states and state parameters is given by p(e/c(l), 0^(1)) for all k in the
frequency grid.
The maximization in (7.36) is made complicated by the 'nuisance' parame-
ters OK{^ T) and t/?, which need to be integrated out. Actually, the piano roll
contains all the information for multiple FO estimation, though the hyperpara-
meters ip may contain information useful for, e.g., instruments classification.
 7 Multiple FO Estimation Based on Generative Models 221

Of course, the integration of the nuisance parameters, as well as the maxi-

mization in (7.36), cannot be performed analytically. Cemgil et al. propose an
algorithm based on message passing, expectation-maximization, and Kalman
filtering on a sliding window which runs along the whole acoustic waveform.
This algorithm jointly estimates the piano roll, the parameters, and the hy-
perparameters. In the special case of single FO estimation, a more efficient
estimation procedure can be implemented. Details about these algorithms
may be found in [78].

7.3.2 Approaches Based on a Sliding Window

The approaches based on noisy sum-of-sines models and sliding windows

make the sliding window Wn(i), located at time n, appear explicitly in the
model. The model in (7.9) becomes, inside the frame located at time n, de-
noted sj^(i) where i is the local time index
J{n) M,

^ni^) = ^ Y 1 otl^{n)wn{i) sin {27rkj^rn{n)i) .r^ ^^.

j=l m=l ^ * ^
+ <^j,m(^)wn(^)cos(27rA:^-,^(n)z) -h en(i),

where the statistics of the noise en(i) are defined framewise. The model in
(7.41) is local in the sense that its parameters have a local interpretation.
In order to fully define the on-line model from a Bayesian viewpoint, the
sequential prior has to be designed.

Approach of Dubois and Davy

Dubois and Davy [158], [160] propose to use Gaussian random walks for both
the frequencies and the amplitude,

k(n) = k ( n - l ) + Vk(n), (7.42)

oc(n) = ot{n - 1) -h Va(n), (7.43)

where k(n) (respectively oc{n)) denotes the vector of frequencies (respectively

amplitudes) for all the partials and notes. The variances of the transition
noises Vk(n) and Va{n) are also assumed to follow a random walk with
log-Gaussian distribution (that is, the logarithm of these variances follows
a Gaussian random walk). The number of partials Mj of note j is set to a
fixed (large) number, and the amplitudes of highest partials is forced to zero
whenever they are above the frequency ks/2. The number of notes is assumed
to be unknown and time varying. A sequential prior is defined as follows:

-fl with probability 1/10

J{n) = J{n-l)-\- { 0 with probability 8/10 , (7.44)
1 with probability 1/10
222 Manuel Davy

which allows the number of notes to increase, decrease, or remain constant.^

Equations (7.41) and (7.42)-(7.44) define a j u m p Markov system [19].
For the sake of simplicity, the frequency evolution model in (7.42) does
not assume links between the frequencies of a given note's partials. Such links
may be built in several ways:

Perfect harmonicity is assumed, and the frequency vector only includes

the fundamental partial frequency. T h e overtone partial frequencies are
computed deterministically from the fundamental partial frequency.
Inharmonicity is allowed, and each partial frequency evolves according
to its own scheme. At the creation of a new note (that is, whenever
J{n) = J ( n l ) - f 1), the new partial frequencies are initialized according to
some probabilistic technique which includes heuristics from deterministic
FO estimation methods.
An inharmonicity model is assumed, for example the piano model, or the
model proposed in [213]. T h e parameters of this model are assumed to
follow a r a n d o m walk, for example. T h e frequency vector only includes
the fundamental partial frequencies.

T h e probabilistic model being defined, parameter estimation can be performed

from the posterior distribution at time n denoted by

p ( k ( l : n ) , a ( l : n ) , J ( l :n),1/^(1 : n ) | x ( l : n ) ) ,

where the notation a ( l : n) is used to denote [ a ( l ) , . . . , a ( n ) ] for any vec-

t o r / s c a l a r a. T h e hyperparameter i/? contains all the model hyperparameters,
such as the time-varying variances of Vk(n) and Vtt(n), the inharmonicity
model parameters, etc.
An interesting feature of this model is t h a t , conditional on J ( l : n ) ,
k ( l : n ) , and i/?(l : n ) , the model is linear and Gaussian: this is because
the model in (7.41) defines a linear and Gaussian likelihood (The noise n{i)
is assumed to be white and Gaussian). From an algorithmic viewpoint, this
permits an efficient estimation strategy based on parallel Kalman filters, each
running with different values of J ( l : n) and '0(1 : n ) . More precisely, it is
possible to implement a Rao-Blackwellized particle filter^ in order to estimate
the frequencies, amplitudes, and number of notes from the posterior distribu-
tion at time n. This algorithm is summarized below, and details may be found
in [158]. T h e overall principle of particle filtering is to run N particles side-by-
side at each time, compute their weights (which are related to their likelihood)
and use these to weight t h e m in the parameter estimation. Particles with low
weights correspond to low-probability parameter successive values, and thus
they are deleted. Particles with high weights have high probability parameter

^The probabilities in (7.44) are indicative, and may be adjusted at will, in par-
ticular when J{n) reaches its minimum/maximum value.
^The reader interested in particle filtering may refer to Chapter 2 for a short
introduction, and to [152] for a full survey.
 7 Multiple FO Estimation Based on Generative Models 223

successive values, and thus they are duplicated in order to better explore the
parameter space around the most probable values.

Algorithm 7.2: Harmonic Tracking Rao-Blackwellized Particle Filter

Initialization
Step 7 . 2 . 1 Initialize the particles at time n = 1.
- For particles j 1,... ,N, sample the initial number of notes J^^\l), the initial
frequency vector k^-^^(l), and the partial amplitudes a^^\l), using some prob-
ability density possibly derived from deterministic estimatioji algorithm applied
to the initial frame 5^. Sample the initial hyperparameter '0^^^(1).
Iterations, for n = 1, 2 , . . .
Step 7 . 2 . 2 The particles are updated.
- For particles j = 1,... ,N, sample thenumbet^of notes using the proposal dis-
tribution q j as follows: J^^\n) - q j ( J ^ ^ ^ ( n ) | j ( ^ ) ( n - l ) , V ^ ( ^ ^ ( n - l ) , x ( l : n ) ) .
- For j = ] ^ , . . . , A/^, sample the frequencies using the proposal distribution qk as
follows: k^^\n) - qi^{k^^\n)\k^^\n - l),J^^\n),ip^^\n - l),x{l : n)).
- For particles j 1 , . . . , TV, update the amplitude vector a^^^ (n) using a Kalman
filter applied to the amplitude evolution model in (7.43) and the likelihood
defined in (7.41), where the other parameters are set to J^^\n), k^^^(n) and

- For j 1 , . . . , A/", updat^ the hyperparameter vector using the proposal dis-
tribution q ^ as follows: \l^^^\n) ~ q^; ('0^"^^(^)|V^^"^H^ ~ 1), a^-^^(n), k^-^^(n),

- For particles j = 1 , . . . , N , compute the importance weight uj^^\n) according

to the rules of importance sampling (see (2.84) in Chapter 2).
Step 7 . 2 . 3 The particles are resampled.
- Duplicate particles with large weights and suppress particles with small weights.
Step 7 . 2 . 4 The parameters are estimated.
- Use the particles in order to compute estimates of the number of notes, the
frequencies, and the related amplitudes.

In Algorithm 7.2, the particles are updated using proposal distributions,

which are designed so as to favour likely parameter values. For example, the
frequencies proposal distribution qk may be build on the current frame fre-
quency spectrum, or from a randomized deterministic FO estimation proce-
dure; see [158]. The performance of the algorithm may also be improved by
adding, at each time n, a particle update step based on one iteration of the
MCMC algorithm presented in Section 7.2.4.

Approach of Vincent and Plumbley

The approach proposed by Vincent and Plumbley [646] uses a model equiva-
lent to that of (7.41), but written in a different form: a cosine and an initial
phase component are used instead of a sine and a cosine, namely
224 Manuel Davy
J{n)M^{n)

= E' ^
j=\ ?n=l
aj,m(n)w(^)cos(27rA;J,m(r^)^ + <^j,m(r^))+e(i), (7.45)

where harmonicity is assumed in each frame, i.e., the partial frequencies are
kj^rn{fT^) mkj^i{n). Similar to the approach by Cemgil et al., it is assumed
that the fundamental frequency belongs to a fixed grid; here, the MIDI semi-
tone scale. The noise probabilistic model is defined in the frequency domain:
roughly, it is Gaussian where the variance in a given auditory frequency band
is proportional to the loudness of s^{i) in that frequency band; see [646].
The parameter estimation procedure is made of two steps, using the
Bayesian setting. First, parameter priors are defined in each frame, indepen-
dently of neighbouring frames. Unknown parameters are estimated framewise
using MAP. Then in the second step, the parameters in different frames are
linked together using sequential priors, and they are re-estimated.
The framewise local priors are defined as follows. Assume k^'iin) is the
frequency in the grid corresponding to the fundamental partial of note j , in
the current frame. The true fundamental partial frequency is assumed to be
close to kjy{n), namely,

p(A;i(n)|fc-r(n)) = CAf{kj.i{ny,k^:['{nlalJ, (7.46)

where A/'(-;-,') is the log-Gaussian distribution.^ In each frame, the num-

ber of partials is set so that Mj{n)kj'[^{n) is below ks/2. The initial phase
prior is uniform; that is, p(0j^^(n)) l^[o,2n]{<Pj,m{'^)) and the ampli-
tudes are assumed to follow a given spectrum envelope denoted Xj^rni'^) for
m = 1 , . . . , Mj{n), Each partial amplitude is assumed to follow a log-Gaussian
distribution
PKm(n)|A,-^(n)) = A / ' K ^ ( n ) ; 7 , ^ ( n ) A , , ^ ( n ) , a i p , ) , (7.47)
where jj{n) is a scale factor with log-Gaussian prior distribution. This com-
pletes the framewise Bayesian model. Such a model is defined for each fre-
quency in the grid, though only a small number of these frequencies are 'active'
in a given frame. Again, similar to Cemgil et al. [78], each grid frequency is
associated to a state ek{n) which may be either 'sound' or 'mute' according
to given prior probabilities. The number of notes J{n) is simply the number
of grid frequencies in the state 'sound'. The framewise Bayesian estimation
procedure consists of finding the state vector exMAp(^) such that
epcMApW = argmax p(ex(n)|s^), (7.48)

where s^ = [5^(1), 5^)^(2),...] and eK{n) is defined as in Section 7.3.1. In

practice, the MAP estimation in (7.48) is based on the joint posterior proba-
bility of the vector e/<:(n) together with the actual frequencies kj^i related to
^'A/^(a; /i, U) is a log-Gaussian distribution' means that log(a) is Gaussian with
mean log(/i) and covariance matrix U.
 7 Multiple FO Estimation Based on Generative Models 225

grid frequencies in the 'sound' state. The other parameters (amplitudes and
7^ factor) have been integrated out using the Laplace approximation method;
see [646]. The actual estimation procedure is based on a greedy, local search
for the states e/c(n), and a gradient-style optimization for the frequencies kj^i.
The note parameters being estimated independently in each frame, they
may now be connected across the frames. Vincent and Plumbley define a se-
quential prior p(efc(n)|eA:(n-l)), and the Viterbi algorithm is used to estimate
the sequence e^MApCl T) from the individual framewise estimates e^MAp(^)*
In other words, a grid frequency can be in the state 'sound' in exMAp(l T) at
time n only if it is already in that state in the local estimate eKMAp(^)- The fi-
nal step consists of re-estimating the actual frequencies and amplitudes from
the posterior p(A:,-i(l : T ) , a , , ^ ( l : ^ ) , 7 , ( 1 : T)\x{l : T ) , e ^ ^ ^ p ( l : T))
that is, given the optimal sequence exuApi^ ^)where the priors in (7.46)
and (7.47) and the prior over 7j (n) have been redefined as log-Gaussian ran-
dom walks.

7.4 O t h e r On-Line Bayesian Approaches

Aside from direct modelling of the acoustic waveform, as presented in the
previous sections, indirect models have been studied. In this section, we review
approaches where the signal spectrogram is modelled, and where harmonic
trajectories are searched.

7.4.1 Methods Designed for Music and General Audio

We present four approaches proposed for general audio processing. All of them
follow harmonic trajectories in the acoustic waveform spectrogram.

Method of Yeh and Robel

The method of Yeh and R5bel [681] applies framewise a model similar to that
of (7.9). This model assumes parallel evolution of the partial amplitudes,
together with the spectral smoothness principle of Klapuri [351] and the
inharmonicity model of Davy and Godsill [122]. The parameter estimation
algorithm is computationally simpler than MCMC, though it is also based on
the generation of candidate notes. Each candidate is evaluated using a score
function, and the best candidates are kept in the final list. This method is
applied framewise.

Spectrogram ModeUing Method of Dubois and Davy

Dubois and Davy [159], [158] extend their method presented in the previous
section to the case where the model is written in the spectrogram domain.
More precisely, the model in (7.41) is changed into
226 Manuel Davy

| jJ{n)
( n ) M^
Mj ^
|DFT.w(.)(/)|^
j=l 771=1

(7.49)
where the noise en{l) is assumed to be zero-mean white Gaussian. The co-
sine term, which was previously used to represent the initial phase, would
be redundant as we consider the power spectrum, and it has been omitted.
The sequential priors are defined as in (7.42)-(7.44). This model is no longer
linear in the amplitudes, and a particle filter close to that in Algorithm 7.2 is
devised; see [158].

Method of Thornburg et al.

Thornburg et al. [625] propose a sequential estimation procedure which uses

the main energy peaks in the spectrogram as input information. It should
be noted that this procedure is aimed at melody extraction, and thus it is
restricted to single FO. The fundamental partial frequency is assumed to be-
long to the MIDI semitone scale, with some unknown, possibly time-varying,
frequency shift from the A4 440 Hz tuning. The amplitudes are also assumed
to belong to a grid with exponential spacing, and the method is designed so
as to associate a state e(n) to each frame s^ at a given time, where e(n)
takes its values in the set {'transient frame onset', 'transient frame contin-
uation', 'pitched frame onset', 'pitched frame continuation', 'silent frame'}.
The transition probability p(e(n)|e(n 1)) is assumed to be given, and only
allows some transitions between the five possible state values. Conditional
on e(n), various transition probabilities are defined for the frequencies, the
amplitudes, and the MIDI scale frequency shift parameter. The unknown pa-
rameters are estimated by MAP Bayesian estimation, using an approximate
forward-backward Viterbi algorithm. The output also includes the sequence
of state e(l : T) which may be used for multiple FO estimation, but also music
segmentation and rhythm quantization.

Method of Sterian et al.

An earlier work is that of Sterian et al. [609]. A Kalman filter was used to
extract sinusoidal partials from the signal. Then, these partials were grouped
into their sources by implementing the grouping principles of Bregman [49] in
terms of individual likelihood functions aimed at evaluating, e.g, the harmonic
concordance. The implementation was based on multiple hypothesis tracking.

7.4.2 Methods for Multiple-Speakers Pitch Tracking

In the speech processing literature, several multiple FO tracking algorithms

have been proposed. A special feature of speech is that the sounds are not
 7 Multiple FO Estimation Based on Generative Models 227

always voiced; that is, portions of a speech signal may not be well modelled
as a noisy sum of sines. Algorithms for multiple-speakers speech tracking take
this feature into account. Moreover, voiced speech is usually quite harmonic,
and thus speech partial frequency models do not incorporate inharmonicity.
Apart from Tabrikian et al. [616] which is for single pitch, Bach and Jordan [25]
and Wu et al. [677] (see Chapter 8) propose multipitch algorithms for speech
processing based on the acoustic waveform spectrogram, or correlogram. The
inference is based on variants of the Viterbi algorithm.
Though developed for a problem different from multiple FO estimation in
musical signals, methods developed for speech yield interesting frameworks
that may be adapted to music.

7.5 Conclusions

In this chapter, several methods for multiple FO estimation have been pre-
sented. Most of them rely on a generative model of the acoustic waveform,
and some model the signal spectrogram. Probabilistic models are defined in
order to make estimation of the generative model parameters feasible.
These methods are quite powerful in the sense that they capture a large
fraction of the information from the acoustic waveform, and this information
may be used for tasks other than multiple FO estimation. Moreover, Bayesian
approaches clearly distinguish model construction from inference (that is, pa-
rameter estimation algorithms). The drawback is that they can be compu-
tationally intensive. However, their computational cost may be reduced by
plugging as many heuristics as possible into the algorithms, for example to
build the proposal distributions in Monte Carlo algorithms, as this makes the
convergence faster. Designing a method of this kind can also be viewed as
a principled way to built algorithms with known theoretical properties (con-
vergence speed, estimation error, etc.). Heuristics can be also used to design
the unknown parameter priors. Finally, generative models may be specifically
designed for various transcription tasks. Models for melody transcription in
complex polyphonic music can surely be based on spectrogram modelling (see
Section 7.4) and are simpler than precise generative models aimed at estimat-
ing the subtle expressive controls of a guitar player.
Work in this vein is in its infancy. More research has to be done to use
as much prior information as possible, and to define more elaborate models,
which could also be used for, e.g., percussion transcription. This will probably
result in quite complex algorithms, but complexity is the main issue in multiple
FO estimation, and it may not be avoided.
8
Auditory Model-Based Methods for Multiple
Fundamental Frequency Estimation

Anssi Klapuri

Institute of Signal Processing, Tampere University of Technology,

Korkeakoulunkatu 1, 33720 Tampere, Finland
Anssi.KlapuriQtut.fi

8.1 Introduction

This chapter describes fundamental frequency (FO) estimation methods that

make use of computational models of human auditory perception and espe-
cially pitch perception. At the present time, the most reliable music tran-
scription system available is the ears and the brain of a trained musician.
Compared with any artificial audio processing tool, the analytical ability of
human hearing is very good for complex mixture signals: in natural acoustic
environments, we are able to perceive the characteristics of several simulta-
neously occurring sounds, including their pitches [49]. It is therefore quite
natural to pursue automatic music transcription and multiple FO estimation
by investigating what happens in the human listener. Here the term multiple
FO estimation means estimating the FOs of several concurrent sounds.
Fundamental frequency is the measurable physical counterpart of pitch.
In Chapter 1, pitch was defined as the perceptual attribute of sounds which
allows them to be ordered on a frequency-related scale extending from low
to high. More exactly, the pitch of a sound was said to be the frequency
of a sine wave that is matched to a target sound by human listeners. The
importance of pitch for hearing in general is indicated by the fact that the
auditory system tries to assign a pitch firequency to almost all kinds of acoustic
signals. Not only sinusoids and periodic signals have a pitch, but even noise
signals of various kinds can be consistently matched with a sinusoid of a certain
frequency. For a steeply lowpass- or highpass- filtered noise signal, for example,
a weak pitch is heard around the spectral edge. Amplitude modulating a
random noise signal causes a pitch perception corresponding to the modulation
frequency. Also, the sounds of bells and vibrating membranes have a pitch,
although their waveform is not clearly periodic and their spectrum does not
have a regular structure. A complete review of this 'zoo of pitch eff^ects' can be
found in [275], [474], [297]. The auditory system seems to be strongly inclined
towards using a single frequency value to summarize certain aspects of sound
230 Anssi Klapuri

events. Computational models of pitch perception attempt to replicate this

phenomenon.
Practical multiple FO estimation methods have a shghtly different purpose
than pitch perception models. The set of acoustic signals of interest is nar-
rower since the physical concept of FO is defined only for periodic and nearly
periodic sounds. Also, the evaluation criteria are different: multiple FO esti-
mation methods are judged based on their reliability in the given task, FO
estimation in a mixture signal, whereas an auditory model should faithfully
reproduce the mechanisms and the behaviour of the auditory system.
For musical sounds, the FO and the perceived pitch are practically equiv-
alent. However, there are ambiguous situations such as the octave ambiguity,
where it is not clear if the FO of a sound is x Hz or half or twice that value.
From the music transcription point of view, it would be desirable to solve
these ambiguities so that the estimated FO would correspond to the perceived
pitch. This is one of the reasons why auditory model-based methods have
been employed. Other reasons include the aim of achieving robustness for
diverse kinds of musical sounds (these are discussed in Section 8.2) and ob-
taining a good time/FO resolution by using a time-frequency decomposition
similar to that in human hearing. The advantages and disadvantages of audi-
tory model-based methods are summarized later in this chapter. In general,
perceptually motivated methods have been quite successful in audio content
analysis.
The primary focus of this chapter is on practical multiple FO estimation
and not so much on auditory modelling. More comprehensive introductions to
pitch perception models can be found in [297], [522], [132]. Also, the emphasis
is laid on multiple FO estimation methods: some perceptually motivated meth-
ods are omitted that are purported to be useful for single FO estimation in
noisy speech signals. The aim of this chapter is twofold: to give a compact de-
scription of pitch perception models so that the reader will be able to develop
auditorily motivated analysis methods of his own and, secondly, to describe
already-existing multiple FO estimators that are based on and motivated by
these models.
This chapter is organized as follows. Section 8.2 discusses the basic acoustic
characteristics of pitched musical sounds and how these can be used to com-
pute the FO of the sounds. Section 8.3 describes computational models of pitch
perception. Section 8.4 introduces music transcription systems which use an
auditorily model as a 'front end'. That is, the systems apply a perceptually-
motivated data representation but the emphasis is laid on the inference that
follows the auditory modelling stage, instead of proposing changes to the
auditory model itself. Section 8.5 describes multiple FO estimation methods
which extend or modify pitch perception models in order to make them bet-
ter applicable to FO estimation in polyphonic music signals. In the end, two
algorithms are described which can be directly used for this purpose. Finally,
Section 8.6 summarizes the main conclusions.
 8 Auditory Model-Based Methods for Multiple FO Estimation 231

10 15 20 25 30 2 3 4 5
Time (ms) Frequency (kHz)

Fig. 8.1. A harmonic sound in the time and frequency domains. The example rep-
resents a vioUn sound with fundamental frequency 290 Hz and fundamental period
3.4 ms.

8.2 Musical Sounds and FO Estimation

This section discusses the acoustic characteristics of pitched musical sounds

and FO estimation when the sounds are presented in isolation. This provides
the background for describing pitch perception models and multiple FO esti-
mation methods in the subsequent sections.

8.2.1 Pitched Musical Sounds

Musical sounds usually consist of several frequency components. The rela-

tive amplitudes of the overtone partials and their time evolution determines
the timbre of the sound. Here we are primarily interested in the frequen-
cies of the partials since FO estimation methods try to normalize away the
timbre information. From this point of view, pitched musical sounds can be
divided into two main classes: sounds that are harmonic and sounds that are
not. The methods to be described in this chapter are concerned with both of
these.
Most Western musical instruments produce harmonic sounds.^ These
sounds have a spectral structure where the dominant frequency components,
called harmonics^ are approximately regularly spaced. Figure 8.1 illustrates a
harmonic sound in the time and frequency domains. The FO of the sound is
the inverse of its time-domain period and the frequency spacing between the
overtone partials corresponds approximately to the FO. Usually the overtone
components are not perceived separately but only the pitch and the timbre of
the entire sound are heard.
For an ideal harmonic sound, the frequencies of the overtone partials are
integer multiples of the FO. However, it should be noted that the spectra
of harmonic sounds are not always perfectly harmonic; the higher-order over-
tones of plucked and struck string instruments deviate slightly from their ideal

^More exactly, all instruments in the chordophone and aerophone families (see
Table 6.1 on p. 167).
232 Anssi Klapuri

i'
| 0
Plllllllf
10 20 30 2000 4000 6000
Time (ms) Frequency (Hz)

Fig. 8.2. A vibraphone sound (FO 330 Hz) illustrated in the time and frequency do-
mains. In the right panel, the frequencies of the most dominant spectral components
are shown in relation to the FO.

harmonic positions. For these classes of instruments, the partial frequencies

obey the formula
fj=jF^l + Bip-l), (8.1)
where F is the fundamental frequency, j = 1,2,... is the partial index, and
B is an inharmonicity factor [193, p. 363]. Typical values of B are of the
order 10~^ or 10~^ for the middle pitch range of the piano, for example. This
makes the higher-order partials gradually shift upwards in frequency, but the
structure of the spectrum is in general very similar to that in Fig. 8.1, and
the sounds can be classified as harmonic. The inharmonicity is due to the
stiffness of real strings, which contributes a restoring force along with the
string tension [193], [315].
Figure 8.2 shows an example of a sound which does not belong to the class
of harmonic sounds although it is nearly periodic in the time domain and has
a clear pitch. In Western music, mallet percussion instruments are a case in
point: these instruments produce pitched sounds which are not harmonic. The
most common instruments in this family are the marimba, the vibraphone, the
xylophone, and the glockenspiel. The sound production mechanism in all of
these is a vibrating bar. A bar of uniform thickness with free ends has vibration
modes whose frequencies are not in integral ratios. However, by making the
bar thinner at the middle of its length, the overtones can be tuned. The first
overtone of the marimba and the vibraphone is typically tuned to be four
times the FO and that of the xylophone to be three times the FO.

8.2.2 Basic Principles of FO Estimation

There are a large number of different methods for monophonic FO estimation

[289]. Comparative evaluations of these can be found e.g. in [535], [290], [134].
The aim of this section is not to make an exhaustive coverage of these, but
merely to point out the main acoustic features that different algorithms are
built upon: time-domain periodicity and frequency-domain periodicity, and to
provide a few representative examples of each approach.
The majority of FO estimation methods are based on measuring the pe-
riodicity of an acoustic signal in the time domain (see e.g. [618], [135]). This
 8 Auditory Model-Based Methods for Multiple FO Estimation 233

makes sense, since all the pitched musical sounds described above are periodic
or almost periodic in the time domain. As reported in [134], quite accurate
single FO estimation can be achieved simply by an appropriate normalization
of the short-time autocorrelation function (ACF), defined as

N-l

n=0
The FO of the signal x{n) can be computed as the inverse of the lag r that
corresponds to the maximum of r{r) within a predefined range. To avoid
detecting an integer multiple of the period, short lags have to be favoured
over longer ones.
An implicit way of measuring time-domain periodicity is to match a har-
monic pattern to the signal in the frequency domain. According to the Fourier
theorem, a periodic signal with period r can be represented with a series of
sinusoidal components at the frequencies j / r , where j is a positive integer.
This can be observed for the musical sounds in Figs. 8.1 and 8.2. Algorithms
that are based on frequency-domain harmonic pattern matching have been
proposed in [153], [54], [428], for example.
Another class of FO estimators measure the periodicity of the Fourier spec-
trum of a sound [384], [380]. These methods are based on the observation that
a harmonic sound has an approximately periodic magnitude spectrum, the pe-
riod of which is the FO. In its simplest form, the autocorrelation function p{m)
over an A^'-length magnitude spectrum is calculated as
N/2-m-l
P('^)-N ^ \X{k)\\X{k + m)\. (8.3)
fc=0

In the above formula, any two frequency components with a certain spectral
interval m support the corresponding FO. The spectrum can be arbitrarily
shifted without affecting the output value. An advantage of this is that the
calculations are somewhat more robust against the imperfect harmonicity of
plucked and struck string instruments since the intervals between the overtone
partials do not vary as much as their absolute frequencies deviate from the
harmonic positions. However, in its pure form this approach has more draw-
backs than advantages. In particular, estimating low FOs is not reliable since
the FO resolution of the method is linear whereas the time-domain ACF leads
to 1/F resolution.
An interesting difference between the FO estimators in (8.2) and (8.3) is
that measuring the periodicity of the time-domain signal is prone to errors
in FO halving because the signal is periodic at twice the fundamental period
too, whereas measuring the periodicity of the magnitude spectrum is prone to
errors in FO doubling because the spectrum is periodic at twice the FO rate,
too. The two approaches can be combined using an auditory model, as will
be described in Section 8.3.2.
234 Anssi Klapuri

Xc(n) Inner hair-cell Zc(n)

to model (neural
B transduction)
Input
o
x(n) T5
3
<

Fig. 8.3. An illustration of the cochlea (left) and its cross-section (middle). The
right panel shows a rough computational model of the cochlea.

8.3 Pitch Perception Models

This section describes computational models of pitch perception and discusses
the advantages that an auditory model-based method may have in multiple
FO estimation.
The human auditory system can be divided into two main parts: peripheral
hearing and the auditory cortex in the brain. Both of these play an important
part in pitch perception. The peripheral part consists of the outer ear, the
middle ear, and the inner ear. The first two of these essentially contribute to
directional hearing and impedance matching of sound. From the pitch analysis
point of view, the interesting part starts from the inner ear, where there is an
organ called the cochlea.
The cochlea is a sophisticated organ where pressure variations are trans-
formed into properly coded neural impulses in the auditory nerve. Physiologi-
cally, the cochlea is a long, coiled, tubular structure which is filled with liquid
and tapers towards its end (see Fig. 8.3). The cochlea is divided into two main
sections by the basilar membrane that runs its entire length. When the me-
chanical vibrations of the eardrum are transmitted via the middle ear to the
inner ear, hydraulic pressure waves are caused in the cochlea and the basilar
membrane starts to vibrate. The waves propagate along the basilar membrane
so that high frequencies peak in amplitude (resonate) near the beginning and
low frequencies get their largest amplitude at the far end.
On the basilar membrane, there is the organ of Corti which contains two
types of hair cells. Outer hair cells are active elements which contribute to the
resolution of the cochlear frequency analysis, making different places along
the basilar membrane more sharply tuned to their characteristic frequencies
than they would be by the acoustic properties of the membrane alone. Inner
hair cells register the movement of the basilar membrane. They respond to
mechanical displacement by generating nerve impulses into the auditory nerve
fibres that are attached to them and lead to the brain [680].
Computational models of the cochlea comprise two main parts which can
be summarized as follows (see Fig. 8.3):
1. An acoustic input signal is passed though a bank of bandpass filters, called
auditory filters, which model the frequency selectivity of the inner ear.
 8 Auditory Model-Based Methods for Multiple FO Estimation 235

Typically about 100 filters are used with centre frequencies uniformly dis-
tributed on a nearly logarithmic frequency scale (details in Section 8.3.1).
The outputs of individual filters simulate the mechanical movement of the
basilar membrane at different points along its length.
2. The signal at each band, or auditory channel^ is processed to model the
transform characteristics of the inner hair cells which produce neural im-
pulses in the auditory nerve. In signal processing terms, this involves three
main characteristics: compression and level adaptation, half-wave rectifi-
cation, and lowpass filtering (details in Section 8.3.2).
In the following, the acoustic input signal is denoted by x{n) and the
impulse response of an auditory filter by gd'f^)^ where c is the channel index.
The output of the auditory filter at channel c is denoted by Xc{n) and functions
as an input to the second step. The output of the inner hair cell model is
denoted by Zc{n) and represents the probability of observing a neural impulse
at channel c.
The processing mechanisms in the brain can be studied only indirectly and
are therefore not as accurately known. Typically the relative merits of differ-
ent models are judged according to their ability to predict the perception of
human listeners for various acoustic stimuli in psychoacoustic tests. Different
theories and models of the central auditory processing will be summarized in
Section 8.3.3, but in all of them, the following two processing steps can be
distinguished:
3. Periodicity analysis of some form takes place for the signals Zc{n) within
the auditory channels. Phase differences between channels become mean-
ingless.
4. Information is integrated across channels.
In the above processing chain, the auditory nerve signal Zc{n) represents
a nice 'interface' between the Steps 2 and 3 and thus between the peripheral
and central processes. The signal in the auditory nerve has been directly
measured in cats and in some other mammals and this is why the stages 1
and 2 are quite well known. Computational models of the peripheral hearing
can approximate the auditory-nerve signal quite accurately, which is a great
advantage since an important part of the processing already takes place at
these stages. However, central processes and especially Step 3 are (arguably)
even more crucial in pitch perception. The above four steps are now described
in more detail.

8.3.1 Cochlear Filterbank

Frequency analysis is an essential part of the cochlear processing. Frequency

components of a complex sound can be perceived separately and are coded
independently in the auditory nerve (in distinct nerve fibres) provided that
their frequency separation is sufficiently large [473]. This frequency analysis
236 Anssi Klapuri

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Frequency (Hz)

1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Frequency (Hz)

Fig. 8.4. Frequency responses of a few auditory filters shown on the logarithmic
{top) and on the linear magnitude scale (bottom). The dashed line in the upper
panel shows the summary response of the filterbank when 70 auditory filters are
distributed between 60 Hz and 7 kHz.

can be modelled with a bank of linear bandpass filters: Figure 8.4 shows an
example of such a filterbank.
The bandwidths and the shape of the power response of the auditory filters
have been studied using the masking phenomenon [192], [499]. Masking refers
to a situation where an audible sound becomes inaudible in the presence of
another, louder sound. In particular, if the distance between two spectral com-
ponents is less than a so-called critical bandwidth^ one easily masks the other.
The situation can be thought of as if the components would go to the same
auditory filter, or to the same channel in the auditory nerve. If the frequency
separation is larger, the components are coded independently and are both
audible.
The bandwidths of the auditory filters can be conveniently expressed us-
ing the equivalent rectangular bandwidth (ERB) concept. The ERB of a fil-
ter is defined as the bandwidth of a perfectly rectangular filter which has a
unity magnitude response in its passband and an integral over the squared
magnitude response which is the same as for the specified filter. The ERB
bandwidths be of the auditory filters have been found to obey

6c = 0.108/c + 24.7Hz, (8.4)

where fc is the centre frequency of the filter at channel c [473].

The centre frequencies of the auditory filters are typically assumed to be
uniformly distributed on a critical-band scale. This frequency-related scale is
derived by integrating the inverse of (8.4), which yields

e(/) = 21.4 logio (0.00437/+ 1). (8.5)

 8 Auditory Model-Based Methods for Multiple FO Estimation 237

In the above expression, / denotes frequency in Hertz and ^(/) gives the
critical-band scale. When / varies between 0 Hz and 20 kHz, ^(/) varies
between 0 and 42. Intuitively, this means that approximately 42 critical bands
(or auditory filters) would fit within the range of hearing if the passbands of
the filters were non-overlapping and rectangular in shape. Conversion from
the critical-band scale back to Hertz units is given by

/ ( O = 229 X (10^/21-^ - 1). (8.6)

For example, let us distribute 70 filters uniformly on the critical-band

scale between 100 Hz and 10 kHz. Using (8.5), we find that the corresponding
frequency boundaries on the critical-band scale are 3.36 and 35.3, respec-
tively, and that the distance between each two centre frequencies has to be
(35.3 3.36)/69 = 0.463 on this scale. The centre frequencies on the critical-
band scale can then be converted to Hertz units using (S.^).
When a lot of auditory filters are uniformly distributed on the scale ^ ( / ) ,
power responses of the filters sum approximately to a flat response, as indi-
cated by the dashed line in Fig. 8.4. Typically about 100 filters are used to
obtain a good sampling of centre frequencies along the cochlea and a suffi-
ciently flat summary response. Note that in this case, the passbands of the
filters overlap considerably. In FO estimation, only the filters up to about 5
to 8 kHz need to be used, as the most significant harmonic components are
below this.
The time-domain impulse responses of the auditory filters have been stud-
ied using a so-called reverse correlation method. In the study by de Boer and
de Jongh [128], the ear of a cat was stimulated with white noise and the
resulting action potentials of individual auditory nerve fibres were recorded
simultaneously. Using the input signal and the recorded train of neural im-
pulses, the impulse response of the corresponding auditory filter was derived.
The impulse response relates the input signal to the firing probability of the
nerve fibre under study, that is, to the probability of an inner hair cell gener-
ating an impulse to the fibre.
A so-called gammatone filter provides an excellent fit to the experimentally
found impulse responses. The filter is defined by its impulse response as [502]

gc{t) = ar-^e-2^^* x COS{2TTfct + (9), (8.7)

where the normalization factor a = (2'Kh)'^/F{n) ensures a unity response at

the centre frequency, r(n) is the gamma function, and the parameter value
n = 4 leads to a shape of the power response that matches best with real
auditory filters. The parameter b = 1.0196c is used to control the bandwidth
of the filter.
Figure 8.5 illustrates the impulse responses of two gammatone filters
with centre frequencies 100 Hz and 1.0 kHz, and with bandwidths obtained
from (8.4). The impulse response consists of a sinusoidal tone at the centre
238 Anssi Klapuri

20 40 10 0 1000 2000
Time (ms) Time (ms) Frequency (Hz)
Fig. 8.5. Impulse responses of two gammatone filters with centre frequencies 100 Hz
(left) and 1.0 kHz (middle). The frequency response of the latter filter is shown on
the right.

frequency of the filter, /c, windowed with a function that is precisely the
gamma distribution from statistics. Frequency responses of several gamma-
tone filters are shown in Fig. 8.4.
The gammatone filters can be implemented efficiently using a cascade of
four second-order IIR filters. A detailed description of the design of the filter-
bank and the corresponding source code can be found in the technical report
by Slaney [591].

8.3.2 Mechanical-to-Neural Transduction

Inner hair cells (IHC) are the elements which convert the mechanical motion of
the basilar membrane into firing activity in the auditory nerve. Each IHC rests
at a certain point along the basilar membrane and thus follows its movement
at this position. Correspondingly, in the computational models the output of
each auditory filter is processed by an IHC model.
The IHCs produce neural impulses, or 'spikes', which are binary events.
However, since there is a large population of the cells, it is conventional to
model the firing probability as a function of the basilar membrane movement.
Thus the input to an IHC model comes from the output of an auditory filter,
Xc{n), and the output of the IHC model represents the time-varying firing
probability denoted by Zc{n).
Several computational models of the IHCs have been proposed. An exten-
sive comparison of eight different models was presented by Hewitt and Meddis
in [291]. In the evaluation, the model of Meddis [456] outperformed the others
by showing only minor discrepancies with the empirical data and by being
also one of the most efficient computationally. An implementation of this
model is available in the AIM [501] and HUTear [273] auditory toolboxes, for
example.
A problem with the realistic IHC models is that they depend critically on
the absolute level of their input signal. The dynamic range of the model of
Meddis [456], for example, is only 25 dB and the firing rate saturates at the
60 dB level. This limitation of individual IHCs is real, and it seems that the
auditory system uses a population of IHCs with different dynamic ranges to
 8 Auditory Model-Based Methods for Multiple FO Estimation 239

achieve the good intensity discrimination performance over a dynamic range

of about 120 dB [523, pp. 137-142]. This has not been included in the com-
putational models of the individual IHCs [291].
For the above-described reason and for the sake of simplicity, many prac-
tical systems have replaced a realistic IHC model by a cascade of (i) compres-
sion, (ii) half-wave rectification, and (iii) lowpass filtering [171], [327], [677],
[354]. As mentioned in the beginning of this section, these are the main char-
acteristics of the IHCs. An advantage of doing this is that the behaviour of
the overall system becomes easier to analyse and the signal-level dependency
is removed. As a disadvantage, the longer-term level adaptation properties of
more realistic IHC models are lost. This is also the approach followed here:
instead of going into the details of realistic IHC models, we analyse the ba-
sic characteristics of the IHC in order to understand their function in pitch
perception and practical FO estimation.
(i) The compression step has taken slightly different forms in different
implementations, but a common theme in all of these has been to scale the
sub-band signals Xc{n) inversely proportional to their variance. Ellis scaled
the variances of the sub-band signals to unity [171]. Klapuri generalized this
approach by scaling the sub-band signals by a factor cr^~^, where ac is the
standard deviation of Xc (n) and 0 < i^ < 1 is a compression coefficient [354].
Tolonen and Karjalainen omitted compression at sub-bands but pre-whitened
the spectrum of an input signal using inverse warped-linear-prediction filter-
ing, which leads to a very similar result [627].
(ii) Half-wave rectification (HWR) is the clearly non-linear processing step
in the mechanical-to-neural transduction. It is defined as

HWR(x)-^ ' - ' (8.8)

As simple as it seems, rectification within the sub-bands plays an important

part in pitch perception and in practical FO estimation. In particular, it allows
a synthesis of the time and the frequency-domain periodicity analysis methods
introduced in (8.2) and (8.3), respectively.
Figure 8.6 illustrates the HWR operation for a narrow-band signal which
consists of five overtones of a harmonic sound. Most importantly, the rec-
tification generates spectral components which correspond to the frequency
intervals between the input partials. The spectral components generated be-
low 1 kHz represent the amplitude envelope of the input signal, as shown
in the lowest panels. A signal that consists of more than one frequency
component exhibits periodic fluctuations, beating, in its time-domain am-
plitude envelope. That is, the partials alternately amplify and cancel each
other out, depending on their phase. The rate of beating caused by each
pair of frequency components depends on their frequency difference and,
for a harmonic sound, the frequency interval corresponding to the FO
dominates.
240 Anssi Klapuri

I
t
10
0

3
1

E
^ -10 5 0
5 10 2000 4000 6000
Time (ms) Frequency (Hz)
10
$ Q UUU/J\JWWU/J\AA/UUIAAAA^
. i 0.5
c
I -10 S 0.
5 10 2000 4000 6000
Time (ms) Frequency (Hz)

1 2 3 0.5
|-2f a
CO

10 S 0 2000 4000 6000

Time (ms) Frequency (Hz)

Fig. 8.6. Upper panels show a signal consisting of the overtone partials 13-17 of a
sound with FO 200 Hz (fundamental period 5 ms) in the time and frequency domains.
Middle panels illustrate the signal after half-wave rectification. Lower panels show
the result of lowpass filtering the rectified signal with a 1 kHz cut-off.

The complex Fourier spectrum Y{k) of a rectified signal y{n) = HWR(x(n))

can be approximated by
N/2-k
Yik) = ^S{k) + lxik) + - J2 X{j)X{k-Jl (8.9)
j=-N/2+k
where 5{k) is the unit impulse function, and X{k) and ax are the complex
Fourier spectrum and the standard deviation of x(n), respectively [353, p. 38],
[117]. The approximation assumes that x{n) is a zero-mean Gaussian random
process but it is sufficiently accurate for signals such as that in Fig. 8.6, too. On
the right-hand side of (8.9), the first term is a dc-component, the second term
represents the spectrum of the input signal, and the last term, the convolution
of the spectrum X{k) with itself, represents the beating components of the
amphtude-envelope spectrum. In addition, the last term generates a harmonic
distortion spectrum centred on twice the centre frequency of the input narrow-
band signal x{n) in Fig. 8.6. Periodicity analysis of the resulting signal in the
time domain (see the next subsection) leads to a combined use of the time
and frequency domain periodicity because the rectified signal consists of both
the input partials and partials that correspond to their diflFerence frequencies.
Another important property of the HWR is that a series of partials with
approximately uniform amplitudes cause strong beating. This is because the
magnitude of beating caused by each two frequency components is deter-
mined by the smaller of the two amplitudes. In the spectrum of a harmonic
sound, each pair of neighbouring harmonics contributes to the beating at the
 8 Auditory Model-Based Methods for Multiple FO Estimation 241

fundamental-frequency rate, but the 'minimum amplitude' property filters

out individual higher-amplitude partials. This phenomenon is well known in
hearing: if the amplitude of one of the overtones of a harmonic sound rises
clearly above the others, it is perceptually segregated and stands out as an
independent sound [49]. In computational multiple FO estimation, this is a de-
sirable characteristic since it makes the FO computations more immune to the
partials of other, co-occurring sounds. Especially when processing the higher
overtones of a sound, this partly prevents stealing the energy of the partials
of other sounds.
(iii) Lowpass filtering the rectified signal can be used to balance the weight
between the amplitude envelope versus the input narrow-band signal. Most
systems have used a fixed low-order lowpass filter with a cut-off frequency
around 1 kHz at all channels. The sub-band signal after compression, rectifi-
cation, and lowpass filtering is denoted by Zc{n).

8.3.3 Periodicity Analysis at Sub-Bands and Cross-Band

Integration

The auditory nerve signal, modelled by Zc{n), c = 1 , . . . , C , is further

processed in the brain. Although the central processing mechanisms are not
accurately known, it has been convincingly shown that periodicity analysis of
some kind takes place within each auditory channel and the results are then
combined across channels to yield a pitch perception [457], [67]. This amount
of knowledge is already very useful and almost carries us to a situation where
only parameter optimization is left in order to process pitch in a way similar
to that of the human brain.
The first pitch model of the above-described type was proposed by Lick-
lider [409]. He proposed to computed short-time autocorrelation functions
'^c{^) within the auditory channels c and to derive pitch from the resulting
two-dimensional {cxr) representation. This became known as the 'duplex the-
ory' of pitch perception because it involved both frequency analysis (by the
cochlear filterbank) and autocorrelation analysis. Further development with
this class of models was made by Lyon [422], Weintraub [663], and Slaney and
Lyon [594].
Meddis and Hewitt implemented Licklider's model using a gammatone
filterbank and a realistic IHC model and carried out extensive simulations to
investigate if the pitch estimate of the model agreed with human listeners for
various audio signals [457]. The authors computed ACFs within the auditory
channels as
n
,(n, T) = ^ Zc{n - i)zc{n - i - r)w{i), (8.10)
i=o
where Zc{n) is the output of the IHC model in channel c and at time
n, rdn^r) is the ACF, and an exponentially decaying window function
w{i) = (l/i7)e~*/^ was applied to give more emphasis to the most recent
242 Anssi Klapuri

samples [457], [459].^ It should be noted that the data structure at this stage
was three dimensional {cxrxn). Across-channel information integration was
then done simply by summing across channels, resulting in a summary ACF

Pitch at time n was estimated by searching the highest peak in s{n^ r) within
a predefined lag range [457, p. 2884].
Meddis and Hewitt demonstrated that the model was able to predict the
perceived pitch for a large set of test stimuli used previously in psychoa-
coustic tests [457]. Moreover, Meddis and O'Mard later noted that the im-
plementation is a special case of a more general model consisting of four
stages: (i) cochlear bandpass filtering, (ii) half-wave rectification and lowpass
filtering, (iii) within-channel periodicity extraction, and (iv) across-channel
aggregation of periodicity estimates [459]. This became known as the unitary
model of pitch perception because the single model was capable of simulating
a wide range of pitch perception phenomena. Different variants of the unitary
model have been used since then in a number of signal analysis systems [171],
[133], [627], [677].
Cariani and Delgutte carried out a direct experiment to find out the char-
acteristics in the auditory nerve signals that correlate with the perceived pitch
[67]. Instead of using a simulated cochlea, the authors studied the signal in
the auditory nerve of a cat in response to complex acoustic waveforms. They
found that the time intervals between neural spikes are particularly important
in encoding pitch. The authors computed histograms of time intervals between
both successive and non-successive impulses in individual auditory nerve fi-
bres, and summed the histograms of 507 fibres to form a pooled histogram.
What the authors noticed was that, for a diverse set of audio signals, the
perceived pitch correlated strongly with the most frequent interspike interval
in the pooled histogram at any given time [67]. This suggests that the pitch of
these signals could result from central auditory processing mechanisms that
analyse interspike interval patterns. Computational models of the cochlea do
not produce discrete neural spikes but rather real-valued signals Zc{n), which
represent the probability of a neural firing (in diff'erent nerve fibres). How-
ever, Cariani and Delgutte noted that the interspike interval codes are closely
related to autocorrelation operations [67, p. 1712]. For a real-valued signal,
ACF can replace the interval histogram.
Despite the above strong evidence, it seems that the ACF is not precisely
the mechanism used for periodicity estimation in the central auditory system,
but some experimental and neurophysiological findings contradict the ACF
(see e.g. [322] and the brief summary in [131, p. 1262]). Meddis and Hewitt,
for example, used the ACF but wanted to 'remain neutral about the exact

^In practice, the windowing and summing can be implemented very efficiently
using a leaky integrator.
 8 Auditory Model-Based Methods for Multiple FO Estimation 243

0 50 100 150 100 200 300 400 500 100 200 300 400 500
Time (ms) Frequency (Hz) Frequency (Hz)

Fig. 8.7. The impulse response {left) and frequency response (middle) of a comb
filter with the feedback delay of 10 ms and feedback gain 0.9. For comparison, the
right panels shows the power response of the ACF for 10 ms lag.

mechanism by which temporal information is extracted from the activity of

the auditory nerve fibres' [457, p. 2879].
A number of alternative mechanisms to the A C F have been proposed [500],
[68], [66], [132]. Although none of these really surpass the modelling power
of the A C F for a large class of signals, comb filter-like solutions have been
proposed by several authors and are therefore discussed in the following. T h e
o u t p u t of a comb filter for an input signal Zc{n) is given by

Vein, r) = {1- a)zc{n) + aydn - r, r ) , (8.12)

where r is the feedback delay and 0 < a < 1 is the feedback gain.
Figure 8.7 shows the impulse response and the frequency response of a
comb filter with a feedback delay r = 10 ms. For comparison, the power re-
ponse of the A C F for t h e corresponding lag r is shown in the rightmost panel."^
As can be seen, the comb filter is more sharply tuned to the harmonic frequen-
cies of the period candidate and no negative weights are applied between these.
Periodicity analysis with comb filters can be accomplished by invoking a
bank of such filters with diflPerent feedback delays r and by computing locally
time-averaged powers at the o u t p u t s of the filters. Figure 8.8 illustrates the
o u t p u t powers of a bank of comb filter for a couple of test signals. In t h e case
of a periodic signal, all comb filters t h a t are in rational-number relations to
the period of the sound show response to it, as seen in panel (b).
A bank of comb filters has been proposed for auditory processing e.g. by
Cariani [66, Eq. (1)], who used the filterbank to separate concurrent vowels
with different FOs. Cariani also proposed a non-linear mechanism which con-
sisted of an array of delay lines, each associated with its characteristic delay
and a non-linear feedback mechanism instead of the linear one in (8.12). Pe-
riodic sounds were reported to be captured by the corresponding delay loop
and thus became segregated from the mixture signal. T h e strobed temporal

^As a non-linear operation, the ACF does not have a frequency response. How-
ever, since the ACF of a time-domain signal is the inverse Fourier transform of its
power spectrum, the power response of the ACF can be depicted for a single period
value.
244 Anssi Klapuri

m
l| a. i d.i
0.8 0.5
0.6
0.4
0.2
o'Li-u ^ '^ Li
1 24 48 72 96 24 48 72 96 24 48 72 96
-0.5

24 48 72 96
X (samples) X (samples) T (samples) X (samples)

Fig. 8.8. Normalized output powers of a bank of comb filters for (a) a sinusoidal
with 24-sample period and (b) an impulse train with the same period. The feedback
delays of the filters are shown on the x-axis and all the feedback gains were 0.9. The
panels (c) and (d) show the ACFs of the same signals, respectively.

integration (STI) mechanism of Patterson [502], [500, p. 186] is closely related

to comb filters too, although the relation is less direct and full details of the
method are beyond the scope of this chapter.
In all the above-described models, the across-band integration has received
a rather small role. For example, Meddis and Hewitt [457] and Cariani and
Delgutte [67] suggest simply summing the autocorrelation functions or period-
icity histograms across channels (see (8.11)). More complex ways of integrating
the information across channels have been proposed, though. These will be
discussed in more detail in Section 8.5.1, in connection with the estimation of
multiple pitches. In particular, a technique called channel selection will be dis-
cussed which attempts to identify the spectro-temporal regions that represent
the target sound and to reject the channels which contain noise or interfer-
ence. Here it suffices to note that the across-channel information integration
takes place in the central auditory system and may thus employ almost any
complex technique. One curious consequence of this is that the pitch of a
sound can be perceived even when two overtone partials of the sound are fed
to the diflFerent ears of a listener [298].

8.4 Using an Auditory Model as a Front End

This section discusses music transcription systems which use an auditory
model as a front end. That is, the systems apply a perceptually motivated
data representation but the emphasis is laid on higher-level processing in-
stead of proposing changes to the auditory model itself. Section 8.5 will dis-
cuss systems which do the latter and, as will be seen, often some practical
modifications are needed in order to make the models more robust in poly-
phonic music signals. However, putting transcription systems under these two
sections primarily serves the purpose of presentation instead of representing
two clear categories.
The intermediate data representations employed between an input signal
and the transcription result are of great importance. An appropriate repre-
sentation facilitates the design of algorithms that use it and often improves
 8 Auditory Model-Based Methods for Multiple FO Estimation 245

Lag = 0 ms Time = 360 ms

Lag (ms) 100 200 300 400 2.1 3.8

Time (ms) Lag (ms)

Fig. 8.9. Illustration of the log-lag correlogram of Ellis [171]. Input signal in this case
was a trumpet sound with FO 260 Hz (fundamental period 3.8 ms). The left panel
illustrates the three-dimensional correlogram volume. The middle panel shows the
zero-lag face of the correlogram which is closely related to the power spectrogram.
The right panel shows one time slice of the volume, from which the summary ACF
can be obtained by summing over frequency.

the analysis result in practice. The idea of using the same data representation
as the human auditory system is therefore very appealing. The aim of this
section is to investigate the advantages and disadvantages of doing this and
to introduce a few auditory-model implementations that have been employed.
For this purpose, three different music transcription systems are briefly intro-
duced. A discussion of other mid-level data representations in acoustic signal
analysis can be found in [173] and in Chapter 3.

8.4.1 Martin's Transcription System

Martin proposed a system for transcribing piano performances of four-voice

Bach chorales [440], [439]. As a front end of his system, Martin used the log-lag
correlogram model of Ellis [171] which is closely related to the unitary model
of Meddis and Hewitt described above. A bank of 40 gammatone filters was
applied, the output of each filter was half-wave rectified and lowpass filtered,
and then subjected to autocorrelation analysis. Specific to Ellis's model is
that the within-channel ACFs are computed only for a set of logarithmically
distributed lag values, 48 lags per an octave. This makes it computationally
feasible to estimate the ACFs continuously over time and not just in discrete
frames. For each lag r and channel c, the signal rc{n,T) = Zc{n)zc{n r) is
computed and then lowpass filtered in the time dimension, analogous to (8.10).
Summary ACFs are obtained by normalizing each ACF by the value at lag
zero and by summing across channels. Figure 8.9 illustrates Ellis's model.
Martin utilized the good time resolution of Ellis's model by tracking sum-
mary ACF peaks through time and by combining temporally continuous peaks
into musical notes. Simple pruning mechanisms were introduced to eliminate
spurious subharmonic peaks in the summary ACF.
246 Anssi Klapuri

The overall system of Martin's was a complex inference architecture (a

blackboard) where knowledge about the spectral structure of harmonic sounds
was combined with rules governing tonal music and with heuristic techniques.
Support for different FOs was sought for in the summary ACF and then com-
bined with the power envelope information to create note hypotheses. Much
of the innovative work was put into developing an extendable software archi-
tecture which allowed the integration of various types of processing modules
to the system.
The first version of Martin's system simply used a time-frequency spectro-
gram as its input [440], but later the author switched to using the auditory
model [439]. Interestingly, Martin mentions a specific reason for switching to
an auditorily motivated data representation: he suspected that the log-lag
correlogram would facilitate the detection of notes in an octave relationship
without introducing explicit instrument models. Although some evidence for
this was presented, no extensive simulations were carried out to support this
conclusion. Also, Martin reported that the correlogram representation indi-
cated chord roots very clearly and that the analysis did not require resolving
individual higher-order harmonic partials in the spectrum [439, p. 10]. Al-
though Martin's transcription system was never formally evaluated, it was
among the first systems to be able to process signals with more than two
simultaneous sounds and thus had a strong influence on subsequent research.

8.4.2 Auditory Scene Analysis Approach of Godsmark and Brown

Godsmark and Brown proposed a system for modelling the auditory scene
analysis (ASA) function in humans, that is, our ability to perceive and recog-
nize individual sound sources in mixture signals [215]. The authors used music
signals as their test material. ASA is usually viewed as a two-stage process
where a mixture signal is first decomposed into time-frequency components
of some kind, and these are then grouped to their respective sound sources.
In humans, the grouping stage has been found to depend on various acoustic
properties of the components, such as their harmonic frequency relationships,
common onset times, or synchronous frequency modulation [49].
Godsmark and Brown used the auditory model of Cooke [100] for the de-
composition stage. This auditory model also uses a bank of gammatone filters
at its first stage. Notable in Cooke's model is that rectification and lowpass
filtering are not applied at the filterbank outputs but only the compression
and level adaptation properties of the IHCs are modelled, amounting to an au-
ditorily motivated bandwise gain control. Thus the overall model can actually
be viewed as a sophisticated way of extracting sinusoidal components from an
input signal, instead of being a complete and realistic model of the auditory
periphery. The frequency of the most prominent sinusoidal component at the
output of each auditory filter is tracked through time using median-smoothed
instantaneous-frequency estimation [100, p. 36] and, in addition, the instan-
taneous amplitudes of the components are calculated. Since the passbands of
 8 Auditory Model-Based Methods for Multiple FO Estimation 247

the gammatone filters overlap, usually several adjacent filters show response
to the same frequency component. This redundancy is removed by combining
the outputs of adjacent channels so as to form 'synchrony strands' which rep-
resent the time-frequency behaviour of dominant spectral components in the
input signal.
The main focus in the work of Godsmark and Brown was on developing a
computational architecture which would facilitate the integration of different
spectral organization (grouping) principles [215]. The synchrony strands were
used as the elementary units that were grouped to sound sources. The authors
reported that these were particularly suitable for modelling the ASA because
the temporal continuity of the strands is made explicit and they are sufficiently
few in number to perform the grouping for every strand.^ Godsmark and
Brown computed various acoustic features for each strand and then performed
grouping according to onset and offset synchrony, time-frequency proximity,
harmonicity, and common frequency movement.
Godsmark and Brown evaluated their model by investigating its ability
to segregate polyphonic music into its constituent melodic lines. This in-
cluded both multiple FO estimation and organization of the resulting notes
into melodic lines according to the applied musical instruments. The latter
task was carried out by computing pitch and timbre proximities between suc-
cessive sounds. Although transcription accuracy as such was not the main
goal, promising results were obtained for musical excerpts with polyphonies
ranging from one to about four simultaneous sounds.

8.4.3 Marolt's Transcriber for Piano Music

Marolt proposed a system for the automatic transcription of piano music [434].
His system was composed of two main parts: a partial tracking module and a
note recognition module. Input to the partial tracking part was provided by
a model of the peripheral hearing where an input signal was passed through
a bank of 200 gammatone filters and the output of each filter was processed
by Meddis's IHC model [456]. Adaptive oscillators were then used to track
partials at the outputs of the IHC models, one oscillator per channel. The
oscillators employed were similar to those proposed by Large and Kolen in
[391], locking their period and phase to the incoming signal. In order to track
harmonically related partials, the oscillators were interconnected to oscillator
nets^ one per each candidate musical note.
Time-delay neural networks (NNs) were trained to recognize musical notes
at the output of the partial tracking module. Each NN was specifically trained
to recognize a certain piano note in its input. The input to the NNs consisted
of the outputs of all the oscillator networks in a few recent time frames and of
the amplitude envelopes at the outputs of the auditory filterbank. Supervised
learning with a large amount of piano music was used to train the NNs.
^Cooke designed his model exactly for this purpose: to support the grouping
activities in ASA [100, p. 14].
248 Anssi Klapuri

Good transcription results were reported for a test set of three real and
three synthesized piano performances. Concerning the use of the auditory
model, Marolt reported that the compression and level adaptation properties
of Meddis's IHC model were important to the system as they reduced the
dynamic range of the signal and thus enabled the system to track small-
amplitude partials.

8.4.4 Summary of Using an Auditory Front End

Specific advantages of using a perceptually motivated data representation were

reported in the above systems. Martin observed that the log-lag correlogram is
a good indicator of chord roots and that the analysis with the model does not
require resolving individual higher-order harmonics, allowing a better time
resolution. Some evidence for detecting two notes in an octave relationship
was presented. Godsmark and Brown reported that the model of Cooke was
particularly suitable for computational ASA since it produced temporally con-
tinuous sinusoidal components which were relatively few in number. Marolt
reported that the dynamic compression and level adaptation properties of
Meddis's IHC model facilitated the use of small-amplitude partials in the
analysis. Finally, an important feature of auditory models that is not explic-
itly mentioned by any of the above authors is that the compression properties
of the IHC models remove timbral information efficiently and thus make the
models more robust for different musical instruments.
The disadvantages of employing an auditory model were not specifically
reported. However, compared to the use of the Fourier spectrum, for example,
it is fair to say that the computational load of an auditory model is signifi-
cantly higher and that the output of the model is not as straightforward to
interpret and understand.

8.5 Computational Multiple FO Estimation Methods

The pitch perception models described in Section 8.3 are not sufficient as such
for accurate multiple FO estimation in real-world music signals. The purpose
of this section is to describe different approaches to extending the models so
that they become applicable in the present task.
The most obvious shortcoming of the pitch perception models is that they
typically account for a single pitch only. Several pitches in a mixture signal
cannot be detected simply by picking several local maxima in the summary
ACF, for example. The models have been tested using very diverse kinds of
acoustic signals but usually not with sound mixtures. Another shortcoming,
related to the first one, is that the models are not robust in polyphonic sig-
nals. Even the global maximum of the summary ACF does not necessarily
correspond to any of the actual pitches in a mixture signal; certain pitch rela-
tionships can confuse the model. In a typical situation, the constituent notes
 8 Auditory Model-Based Methods for Multiple FO Estimation 249

of a musical chord match the overtones of a non-existing chord root and the
highest peak in the summary ACF indicates the chord root instead of one
of the component sounds.^ Further, the pitch models do not address robust-
ness against additive noise: drum sounds often accompany the pitched sounds
in music. Finally, the computational complexity of the models is rather high
since they involve periodicity analysis at a large number of sub-bands.
On the other hand, there are several issues that are quite efficiently dealt
with using a pitch model. These were summarized in Section 8.4.4 above.
In the following, a number of different methods are described that aim at
overcoming the above-mentioned shortcomings. Some of these were designed
for two-speaker speech signals but are included here in order to cover the
substantial amount of work done in the analysis of multiple-speaker speech
signals. This is followed by a more detailed description of two multiple FO
estimation methods for music signals. It should be noted that the main interest
in this section is not to model hearing but to address the practical task of
multiple FO estimation.

8.5.1 Multiple FO Estimation in Speech Signals

Multiple FO estimation is closely related to sound separation. An algorithm

that is able to estimate the FO of a sound in the presence of other sounds is, in
effect, also assigning the respective spectral components to their sound sources
[49, p. 240]. Separation of speech from interfering speech for the purpose of
its automatic recognition is an important area of sound separation. Here we
look at methods that have utilized pitch information to carry out this task. A
couple of state-of-the-art methods are described, with the aim of discussing
the basic mechanisms that have been used to extend an auditory model to
process multiple pitches.
Multiple FO estimation in speech signals is in many ways a more con-
strained task than in music: the FO range is limited to about three octaves
and the described methods attempt to estimate only two simultaneous FO
tracks. However, the described basic mechanisms are not restricted to speech
signals, and many of them can be generalized to the case of more than two
simultaneous sounds.

Channel Selection

Meddis and Hewitt extended their pitch model (see p. 241) to simulate the
human ability to identify two concurrent vowels with different FOs [458]. The
proposed method included a template-matching process to recognize the vow-
els too, but here only the FO estimation part is summarized. It consists of the
following steps:

^Examples of such chords are the major triad and the interval of a perfect fifth.
250 Anssi Klapuri

1. The pitch model of Meddis and Hewitt is appUed [457]. This involves
a bank of gammatone filters, Meddis's IHC simulation, within-channel
ACF computation, and across-channel summing. The highest peak in the
summary ACF within a predefined lag range is used to estimate the FO
of the more dominant sound.
2. Individual channel ACFs that show a peak at the period of the first de-
tected FO are removed. If more than 80% of the channels get removed,
only one FO is judged to be present and the algorithm terminates.
3. The ACFs of the remaining channels are combined into a new summary
ACF from which the FO of the other vowel is derived.
The authors did not give statistics on the FO estimation accuracy, but
reported clear improvements in vowel recognition as the FO difference of the
two sounds was increased from zero to one semitone or beyond.

Time-Domain Cancellation

The above channel selection scheme can be seen as an instance of a more

general iterative approach where FO estimation is followed by the cancellation
of the detected sound from the mixture, and the estimation is then repeated
for the residual signal. This generalization was pointed out by de Cheveigne,
who further proposed that the cancellation can take place in the time domain
[129], [130]. When the period TQ of one sound in the mixture has been found,
the sound can be removed by applying a cancellation filter with the impulse
response
hroin) = S{n) - S{n - TQ), (8.13)
where 6{n) is the unit impulse function. Convolving an input signal x(r) with
hroin) yields hro{n) 0 x{n) = x{n) x{n TQ) and, if the detected sound is
perfectly periodic, the above filter completely removes it from the mixture. As
a side-effect, however, the filter also removes the partials of other sounds that
coincide with those of the sound being cancelled. Also, a more sophisticated
filter is needed to cancel a sound whose period is not precisely a multiple of
the sampling interval [382].
An advantage of the time-domain cancellation is that it is not bound to
the resolution of the cochlear filterbank and, in principle, it works even when
all the channels are dominated by a single period. The filtering can be done
directly for the input signal or within the channels of an auditory model.
These two are equivalent unless the within-channel filtering is done after the
non-linear IHC simulation stage.
De Cheveigne used the cancellation principle for the actual FO estimation,
too. He proposed to calculate a squared difference function (SDF) which is
defined for an input signal x{n) as
N-l
SDF(n, r)=Y^ {x{n - i) - x{n - i - r ) ) ^ (8.14)
1=0
 8 Auditory Model-Based Methods for Multiple FO Estimation 251

where N is the analysis frame size [131].^ By expanding the square, it can
be seen that SDF(n,r) = E{n) -h E{n r) 2r(n,T), where E{n) denotes
the signal power at time n and r(n,T) is the ACF. Thus the SDF and the
ACF are functionally equivalent, and period estimation can be carried out by
searching for minima in the SDF instead of maxima in the ACF. De Cheveigne
also proposed a joint cancellation model, where two cancellation filters with
periods TA and TB were applied in a cascade so as to cancel two periodic
sounds. By computing the power of the resulting signal as a function of the two
periods, the FOs were found by locating the minumum of the two-dimensional
function [129], [133].
De Cheveigne evaluated both the iterative and the joint FO estimation
method for mixtures of two-voiced speech segments [129]. The iterative algo-
rithm was reported to produce estimates which were correct within 3% accu-
racy in 86% of the frames and the exhaustive joint estimator produced correct
estimates in 90% of the frames. Computational complexity is a drawback of
the joint estimator.

Channel and Peak Selection

Wu, Wang, and Brown proposed an algorithm for tracking the FOs of two si-
multaneous speakers, taking particular interest in noise robustness [677]. Their
method employed a computational model of the peripheral auditory system,
after which the channels significantly corrupted by noise were excluded. From
the remaining channels, ACF peaks were selected so that peaks judged to
give misleading information were rejected. This led to an intermediate data
representation which consisted of only the lag values and channel labels of the
selected ACF peaks (discarding peak amplitudes). The information was then
processed using statistical models.
In more detail, the channel and peak selection process was the follow-
ing. First, a gammatone filterbank was applied and the resulting channels
were classified as 'low-frequency' or 'high-frequency' channels depending
on whether their centre frequency was below or above 800 Hz. Normalized
ACFs were then computed for the low-channel signals directly and for the am-
plitude envelopes of the high-channel signals. Low channels were selected (i.e.,
included in further computations) if the highest peak of the normalized ACF
exceeded a given threshold value. High-frequency channels were selected if the
shapes of the normalized ACFs computed in 16 ms and in 32 ms frames were
sufficiently similar. Peak selection, in turn, consisted of two main rules. First,
an acceptable peak (peak not due to noise) was required to show a submulti-
ple peak at twice its lag value. At high-frequency channels, envelope beating
at the FO rate was assumed and, therefore, subharmonics of any peak higher
than a threshold value were removed. Full details can be found in [677].
^The SDF is closely related to the average magnitude difference function
(AMDF) that has been used to estimate the FO of speech [549]. The AMDF is
obtained by summing absolute values instead of their squares in (8.14).
252 Anssi Klapuri

Input
Highpass Half-wave rectify Periodicity
_1_
Pre-whitening
at 1 kHz Lowpass at 1 kHz detection Summary ACF
Lowpass Periodicity > ^ enhancer
at 1 kHz detection
Enhanced
summary ACF

Fig. 8.10. Block diagram of the pitch analysis method proposed by Karjalainen
and Tolonen [627]. 2005 IEEE, reproduced here by permission.

The remaining channels and peaks were subjected to statistical modelling.

Using clean speech as training material, the difference 6c = TC TQ between a
true (annotated) fundamental period TQ and the period of the closest selected
peak Tc at channel c was studied. The statistical distribution of Sc was used to
determine the likelihood of the observed peaks at channel c given a fundamen-
tal period candidate r. Different observation likelihood functions were defined
for the cases of zero, one, and two FOs (two FOs were jointly estimated). Fi-
nally, a hidden Markov model was employed to model the dynamic aspects of
the FO contours. This included both the continuity of the FO tracks and jump
probabilities between the state spaces of zero, one, or two FOs.
In evaluations, Wu et al. used ten voiced utterances to generate mixtures
of two voices. These were mixed with realistic noise signals, including har-
monic interference and interfering speech signals. Five utterances were used
for training and five for testing. Good results were reported for this database
and an implementation of the method is publicly available [677].

8.5.2 Multiple FO Estimator of Karjalainen and Tolonen

Karjalainen and Tolonen proposed a computationally efficient version of the

unitary pitch model (see p. 241) and extended it to the multiple FO estima-
tion of musical sounds. [327], [627] Figure 8.10 shows the block diagram of
their method. The most obvious difference from the original auditory model
is that the method divides an input signal into two channels only, below
and above 1 kHz, and then analyses the periodicity of the low-channel signal
and of the envelope of the high-channel signal. Despite the drastic reduction
in computation load compared to the unitary pitch model, many important
characteristics of the model were preserved.
The method included several features to address practical robustness is-
sues. Robustness against timbral variation (different musical instruments for
example) was achieved by pre-whitening the input signal using inverse warped-
linear-prediction filtering [272]. In essence, this fiattens the spectral energy
distribution but does not affect the spectral fine structure.
Periodicity analysis in the method of Karjalainen and Tolonen was carried
out using a generalized ACF, originally proposed by Indefrey et al. in [306].
According to the Wiener-Khintchine theorem, the ACF of a time-domain
 8 Auditory Model-Based Methods for Multiple FO Estimation 253

signal X is the inverse Fourier transform of its power spectrum [276, p. 334].
The generalized ACF, then, is defined as

r{r) = IDFT(|DFT(x)r), (8.15)

where DFT and IDFT denote the discrete Fourier transform and its inverse,
and a is a free parameter which determines the frequency domain compres-
sion.^ The standard ACF is obtained by substituting a = 2. Definition of
the cepstrum of x is analogous to ACF and is obtained by replacing the sec-
ond power with the logarithm function. The difference between the ACF and
cepstrum-based FO estimators is quantitative: raising the magnitude spectrum
to the second power emphasizes spectral peaks in relation to noise but, on
the other hand, further aggravates spectral peculiarities of the target sound.
Applying the logarithm function causes the opposite for both. And indeed,
ACF-based FO estimators have been reported to be relatively noise immune
but sensitive to formant structures in speech, and vice versa for cepstrum-
based methods [535]. As a trade-off, Karjalainen and Tolonen suggested using
the value a = 0.67.
Extension to multiple FO estimation was achieved by cancelling subhar-
monics in the summary ACF (SACF) by cHpping the SACF to positive values,
time-scaling it to twice its length, and by subtracting the result from the orig-
inal clipped SACF. This cancellation operation was repeated for time-scaling
factors up to about five. From the resulting enhanced SACF, all FOs were
picked without iterative estimation and cancellation. In more detail, the en-
hancing procedure was as follows:

Algorithm 8.1: Enhancing Procedure of Karjalainen and Tolonen

1. The enhanced SACF s{r) Is Initialized to be equal to the SACF S(T). The scaling
factor m is initialized to value 2.
2. The original SACF is time-scaled to m times its length and the result is denoted by
Smir). Using linear Interpolation,

Sm{T) = S{d) + ^ ^ ^ {S{d + 1) - S{d)) , (8.16)

m
where d = [ r / m j and [J denotes rounding towards negative Infinity.

3. The enhanced SACF is updated as

s(r) ^ max(0, s{r) - max(0, Sm{r))). (8.17)

4. Increment m by 1. If m is smaller than 6, return to Step 2.

The above enhancing procedure is surprisingly efficient in removing spu-

rious peaks from the SACF and in reveahng more than one FO in it. Also,
^In practice, the analysis frame x has to be zero-padded to twice its length before
the first transform.
254 Anssi Klapuri

High-channel ACF ESACF ESACF

IAJLAJULXXJ SACF
SACF
Low-channel ACF

0 5 10 15 0 5 10 15 5 10 15
Lag (ms) Lag (ms) Lag (ms)

Fig. 8.11. Left: The ACFs at the low and the high channel for a viohn sound (FO
523 Hz). Middle: SACF and enhanced SACF for the same sound. Right: SACF and
enhanced SACF for a major triad chord played by the trumpet (FOs 220 Hz, 277 Hz,
and 330 Hz). The circles indicate the correct fundamental periods.

it partly solves the 'chord root' problem mentioned in the beginning of Sec-
tion 8.5 since the enhancing procedure scales the true FO peaks to the position
of the chord root and, if a note does not truly appear at the root, the spu-
rious peak becomes cancelled. The only place where care has to be taken is
in setting values of the original SACF to zero in the lag range [0, /s/1000 Hz]
before the enhancing (here /s denotes the sampling rate). This ensures that
the values on the r = 0 hill do not spread and wipe away important infor-
mation. Zeroing the mentioned lags causes no harm for the analysis since the
algorithm cannot detect FOs above 1 kHz.
Figure 8.11 illustrates the enhancing procedure for an isolated sound and
for a musical chord. As mentioned by Martin [439], the SACF indicates the
non-existing FO of the chord root in the latter case. After enhancing, however,
the true FOs are revealed.
Overall, the method of Karjalainen and Tolonen is quite accurate and it
has been described in sufficient detail to be exactly implementable based on
[627] and on the Matlab toolbox for frequency-warped signal processing by
Harma et al. [272]. A drawback of the method as stated by the authors is
that it is 'not capable of simulating the spectral pitch' [627, p. 713], i.e., the
pitch of a sound whose first few harmonics are above 1 kHz. In practice, the
method is most accurate for FOs below about 600 Hz. Later, Karjalainen and
Tolonen also proposed an iterative approach to multiple FO estimation using
the described simplified auditory model [328].

8.5.3 Multiple FO Estimator of Klapuri

Klapuri's multiple FO estimator for music signals was originally described

in [353, Ch. 4] and later improved and simpified in [354]. The method con-
sists of a model of the peripheral auditory system followed by a periodic-
ity analysis mechanism where FOs are iteratively estimated and cancelled
(Fig. 8.12).
 8 Auditory Model-Based Methods for Multiple FO Estimation 255

Zc(n)
UoCk) Cumulate
Xc(n) Compress, iZe(k)i y Ud(k)
c
CO IDFT(.)I ^
rectify, T
\ U(k) ^
Input Periodicity Normalize,
jfK+j vV ^ detection peak picking

o
<
_u
/"^ Wk)
Fundamental
periods x ,r

Fig. 8.12. Block diagram of the multiple FO estimator of Klapuri [354]. 2005
IEEE, reproduced here by permission.

Model of the Peripheral Auditory System

In the peripheral hearing model, an input signal was first passed through a
bank of gammatone filters with centre frequencies uniformly distributed on
the critical-band scale (see (8.5)) between 60 Hz and 5.2 kHz. A total of 72
filters were employed using the implementation of Slaney [591].
Hair cell transduction was modelled by compressing, half-wave rectifying,
and lowpass filtering the sub-band signals. The compression was implemented
by simulating the full-wave i/th law compression (FWC), which is defined as

x^, X > 0,
FWC(x) = (8.18)
-{-xY, x<0.
For a narrow-band signal, such as the output of an auditory filter, the effect
of the FWC within the passband of the filter can be accurately modelled by
simply scaling the signal with a factor

7c = a{ac) ly-l (8.19)

where ac is the standard deviation of the signal at channel c and the scalar
a depends on u but is common to all channels and can thus be omitted [353,
p. 37]. In addition to the scaling mentioned, FWC generates small-amplitude
distortion components at odd multiples of the channel centre frequency. These
were avoided by using the model (8.19) instead of (8.18) directly.
The FWC provides a single parameter u which determines the degree of
spectral whitening applied on an input signal. The scaling factors 7c normalize
the variances of the sub-band signals towards unity when 0 < i/ < 1. Here,
the value i/ = 0.33 was applied.
The compressed sub-band signals were half-wave rectified by constraining
negative values to zero. As shown in Fig. 8.6, this generates spectral com-
ponents near zero frequency and on twice the channel centre frequency. The
rectified signal at each channel was steeply lowpass filtered with a cut-off
frequency 1.5 times the channel centre frequency in order to attenuate the
distortion spectrum at twice the centre frequency but to pass the sub-band
signal along with its amplitude envelope spectrum. The rectified and lowpass
filtered signals Zc{n) were then subjected to periodicity analysis.
256 Anssi Klapuri

Periodicity Analysis

The periodicity analysis mechanism proposed by Klapuri is best understood

by comparing it with the ACF-based method employed by Meddis and Hewitt
(see p. 241). Short-time ACF estimates within the channels can be efficiently
computed as Vc^ni^) IDFT(|Zc,n(^)P), where IDFT denotes the inverse
Fourier transform and Zc^n{k) is the Fourier transform of Zc{n) computed in
a time frame that is centred at time n and zero-padded to twice its length
before the transform. The within-band ACFs are then summed to obtain the
summary ACF, Sn{r) = Ec^c,n(r).
Because the IDFT and the summing are linear operations, their order
can be reversed and we can write Sn{r) = lDFT{Sn{k)), where Sn{k) =
Ylc l^c,n(^)P- The spectra of real-valued (audio) signals are conjugate sym-
metric and the IDFT can therefore be written out as
K/2-1 . .
s(r) = IDFT(5(fc)) = - | J] cos ( - ^ J 5(fc), (8.20)

where K is the length of the transform frame after the zero-padding.

Klapuri made three modifications to (8.20). First, as seen in Fig. 8.12,
magnitude spectra were summed across channels instead of power spectra.
Analogous to the generahzed ACF in (8.15), it was observed that raising the
magnitude spectra to the second power accentuates timbral peculiarities that
cannot be completely removed by band wise compression in polyphonic signals.
Therefore, within-band magnitude spectra were summed to obtain a summary
magnitude spectrum (SMS),

U{k) = Y,\Zc{k)\, (8.21)

C

where the time index n has been omitted to simplify the notation in the
following. The SMS functioned as an intermediate data representation and all
the subsequent processing took place using it only.
Figure 8.13 illustrates the bandwise magnitude spectra |Zc(A:)| for a saxo-
phone sound. As can be seen, the within-channel rectification maps the contri-
bution of higher-order partials to the position of the FO and its few multiples
in the spectrum. Most importantly, the degree to which an individual overtone
partial j is mapped to the position of the fundamental increases as a function
of j . This is because the auditory filters become wider at higher frequencies
and the partials thus have larger-magnitude neighbours with which to gener-
ate the difference frequencies (beating) in the envelope spectrum. Klapuri's
method was largely based on this observation, as will be explained below.
The second modification concerned the function cos(-) in (8.20), which can
be seen as a harmonic template that picks overtone partials of the frequency
K/T in the spectrum (see the rightmost panel of Fig. 8.7 on p. 243). The func-
tion was replaced by a response that is more sharply tuned to the frequencies
 8 Auditory Model-Based Methods for Multiple FO Estimation 257
' ' ' 1 1 ' ' I . . 1 . 1 1 1 1 1 . 1 1 1
A 1 . ....111.
^4982
A ...ll....
:i: 3680 A . ,.11.
^2703 A . . 1 1
g 1970 . .1 .
^ 1421 . 1
g' 1009 1 .
' ^ 699 1
g 467 1
293 ft
O 163
65 r=. r1 1 . , 1 , , , n1 , r-,-^ ,1 . . , . lH .

100 200 500 1000 2000 5000

Frequency (Hz)

Fig. 8.13. The spectra |-^c(/c)| at a few channels for a tenor saxophone sound (FO
131 Hz).

of the harmonic overtones of a FO candidate and employs no negative weights

between the partials. In practice, the frequency response resembled that of
a comb filter shown in Fig. 8.7. This modification alleviates the interference
of other, co-occurring sounds. Moreover, instead of pointwise multiplying the
complete spectrum U{k) with a comb filter response and then summing, it
was found sufficient to sum up spectral components near the positions of the
peaks of the comb-filter response (see (8.22) below). This led to a very effi-
cient implementation computationally and is closely related to the harmonic
selection methods reviewed by de Cheveigne in [129], and to the harmonic
transform of Walmsley et al. [657].
The relative strength, or salience, A(r) of a fundamental period candidate
r was calculated in Klapuri's system as

A(r; md.x[HLp{k)U{k)] (8.22)

j=i

where /s denotes the sampling rate and the factors / s / r and Hi,p{k) are
related to the third modification to be explained later. The set Kj^r defines a
narrow range of frequency bins in the vicinity of the jth overtone partial of
the FO candidate / s / r . More exactly, Kj^r = [k^^l.kj^l], where
1,(0)
[jK/{T + AT/2)\+h (8.23)

Tn^x{[jK/{T-AT/2)\M,h (8.24)

In the above formulas, K is the transform length and the scalar AT = 1

denotes spacing between successive period candidates r. A uniform sampling
of lag values was used, analogous to the ACF. Equations (8.23)-(8.24) define
the sets Kj^r so that, for a fixed partial index j , all the spectral components
belong to the range of at least one period candidate r, and the ranges of
adjacent period candidates cannot overlap by more than one frequency bin.
258 Anssi Klapuri

The third modification in (8.22) compared to (8.20) is that individual

partials in the sum in (8.22) are weighted by / s / r x Hi,p{k), where the lowpass
response is

^ - ( ^ ^ = 0.108/3fe/V + 24.7- ^'-''^

By comparison with (8.4), it is easy to notice that this is the reciprocal of
the bandwidth of an auditory filter centred at frequency bin k. The factor
/ s / r X Hi,p{k) can therefore be written as F{T)/bc{jF{T)), where F{T) = / s / r
is the FO of the period candidate r (i.e., the frequency interval between its
overtones) and bc{jF{r)) is the width of an auditory filter centred at its j t h
overtone. The ratio of these two was interpreted as the resolvability of the
partial j [353, p. 45]. The lower-order overtones of a harmonic sound are
resolved into separate auditory channels, whereas the higher-order overtones
go to the same auditory channel with their neighbours and their frequencies
cannot be perceived separately (resolved). Actually the lowpass filter Hi,p{k)
would belong to the within-band IHC modelling stage but, since the filter
is the same for all channels, it is equivalent to apply it after the channels
have been combined. The higher the centre frequency of an auditory channel,
the more the filter attenuates the spectrum at the passband of the auditory
filter and thus gives it a smaller weight in relation to the envelope spectrum,
which is around zero frequency and not much affected. This corresponds to
the fact that, at higher auditory channels, the neural firing activity more
and more follows the amplitude envelope of the sub-band signal and not its
fine structurethis is directly related to the concept of resolvability. Discrete
categorization into 'low' and 'high' channels is not needed.
The degree of resolvability as modelled above (and thus the weight of a
partial in the sum in (8.22)) is approximately inversely proportional to the
harmonic index j when r is fixed. As a consequence, the sum in (8.22) can be
limited to j 20 since weights beyond this are relatively small.
Taken together, the computation of the salience function A(r) can be seen
as a process where partials are picked from harmonic positions of the spectrum
U{k), their magnitudes are weighted by the estimated resolvability / s / r x
-f^Lp(fc), and then summed. What makes all the difference is that the within-
channel rectification maps the contribution of higher-order partials to the
position of the fundamental and its few multiples in the spectra Zc{k), and
the degree to which an individual overtone partial j is mapped to the position
of the fundamental increases as a function of j , as explained above. As a
consequence, the whole harmonic series of a sound contributes to its salience,
despite the weighting with resolvability.
The above-described benefit of bandwise rectification cannot be overem-
phasized. Assigning the higher-order partials to their respective sound sources
in polyphonic music signals is a nightmare. The rectification operation ac-
complishes this 'automatically' by mapping the support from higher-order
harmonics to the position of FO and its few multiples in U{k). Figure 8.14
 8 Auditory Model-Based Methods for Multiple FO Estimation 259

^ ^
& & 1 11..
0.1 0.2 0.5 1 2 0.1 0.2 0.5 1 2 5
Frequency (kHz) Frequency (kHz)

3 g
'WA.-A-AJ^ VjuvJl^^ '''^-^^~-'-AJ^-A>vA,^A^^,A>A-JL^AAJL^^^

0.1 0.2 0.5 1 2 0.1 0.2 0.5 1 2 5

Frequency (kHz) Frequency (kHz)

Fig. 8.14. The upper panels show the summary magnitude spectrum U{k) for a
saxophone sound with FO 131 Hz {left) and a violin sound with FO 1050 Hz (right).
The lower panels show the corresponding salience functions A(r).

illustrates the calculation of A(r) for the saxophone sound shown in Fig. 8.13,
and for a violin sound with the FO 1050 Hz.

Iterative Estimation and Cancellation

The global maximum of the function A(T) was found to be a robust indicator
of one of the correct FOs in polyphonic signals. As with most FO estimators,
however, the next-highest salience was often assigned to half or twice that of
the first detected FO. Similarly to de Cheveigne (see p. 250), Klapuri employed
an iterative technique where FO estimation was followed by the cancellation
of the detected sound from the mixture and the estimation was then repeated
for the residual signal. Algorithm 8.2 summarizes the applied technique [354].

Algorithm 8.2: Multiple FO Estimator of Klapuri

1. A residual SMS (7R(/C) is initialized to be equal to U{k). A summary spectrum of
all detected sounds, Uuik), is initialized to zero.
2. A fundamental period f is estimated using UR{k) and (8.22).
3. Harmonic selection is carried out for the found period f according to (8.22)-(8.24).
However, instead of summing up the magnitude values, the precise frequency and
amplitude of each partial is estimated and used to calculate its magnitude spectrum
at the few surrounding frequency bins.
4. The magnitude spectrum of the j t h partial is weighted by fs/TxHhp{kj) and added
to the corresponding position of UD{k) which represents the cumulative spectrum
of all the detected sounds.
5. The residual SMS is recalculated as

Unik) ^ max(0, U{k) - dUuik)), (8.26)

where d = 0.5 controls the amount of the subtraction and is a free parameter of
the algorithm.
6. Return to Step 2.
260 Anssi Klapuri

An important characteristic of the Step 4 is that, before adding the partials

of a detected sound to C/D(A:), they are weighted by their resolvabihty in the
same manner as at the FO detection stage. As a consequence, the higher-
order partials are not entirely removed from the mixture spectrum when the
residual UR{k) is formed. This principle is important in order not to corrupt
the sounds that remain in the residual and have to be detected at the coming
iterations. The described weighting limits the effect of the cancellation to the
lowest harmonics but, as explained above, the higher-order harmonics have
been mapped to the position of the fundamental by the rectification and
are thus effectively cancelled, too.

8.5.4 Results

Simulation experiments were carried out to evaluate the performance of the

method of Tolonen and Karjalainen [627] and that of Klapuri [354]. Implemen-
tations of the method of Wu et al. [677] and Marolt [434] are publicly available
too, but these would have required a specific experimental setup since the for-
mer was designed to process continuous two-speaker speech signals and the
latter to transcribe piano music only.
The acoustic material consisted of samples from the McGill University
Master Samples collection [487], the University of Iowa website,^ IRC AM
Studio Onhne,^ and of independent recordings for the acoustic guitar. There
were altogether 32 different musical instruments, comprising brass and reed
instruments, strings, flutes, the piano, the guitar, and mallet percussion in-
struments. The total number of samples (individual notes) was 2842.
Semi-random sound mixtures were generated by first allotting an instru-
ment and then a random note from its playing range. This was repeated to get
the desired number of simultaneous sounds, which were then mixed with equal
mean-square levels. One thousand test cases were generated for mixtures of
one, two, four, and six sounds.
One analysis frame immediately after the onset ^ of the sounds was fed to
the multiple FO method. The number of FOs to extract, i.e., the polyphony,
was given along with the mixture signal. A correct FO estimate was defined
to deviate less than 3% from the nominal FO of the sound, making it round
to a correct note on the Western musical scale. Two different error rates were
computed. Multiple FO estimation error rate was defined as the percentage of
all FOs that were not correctly detected in the input signals. In predominant
FO estimation, only one FO in the mixture was being estimated and it was
defined to be correct if it matched the correct FO of any of the component
sounds.

^University of Iowa samples: theremin.music.uiowa.edu/MIS.html

^IRCAM Studio Online: soleil.ircam.fr
^The onset of the sounds was defined to be at the point where the waveform
reached one third of its maximum value during the first 200 ms of its playing.
 8 Auditory Model-Based Methods for Multiple FO Estimation 261

Reference
50
4 6 ms - I

M
40 F0<2.1kHzl I
30
20
10
0
1 2 4 6
Polyphony

Reference
93 ms
I 30 F0<2.1kH2

1 2 4 6
Polyphony
1 2 4 6
Polyphony
20
^ 10
5
0 jjl
1 2
li
4 6
Polyphony

Fig. 8.15. FO estimation error rates as a function of the number of concurrent

sounds (polyphony) for the method of Tolonen and Karjalainen [627], the method
of Klapuri [354], and the reference method [351]. The black bars and the white bars
show the multiple FO and the predominant FO estimation error rates, respectively.
The upper panels show the results for a 46 ms analysis frame and the lower panels
for a 93 ms frame.

The left-hand panels of Fig. 8.15 show the error rates for the method of
Tolonen and Karjalainen in 46 ms and 93 ms analysis frames. The FO range in
these experiments was limited to the three octaves between 65 Hz and 520 Hz,
because the accuracy of the method was found to degrade rapidly above 600 Hz
(see Section 8.5.2). The black bars show the multiple FO estimation error
rates and the white bars show the predominant FO estimation error rates.
The global maximum of the enhanced SACF was used for the latter purpose.
The method performed robustly in polyphonic mixtures, and especially the
predominant FO estimation error rates remained reasonably low even in short
time frames and in rich polyphonies. Taking into account the computational
efficiency (faster than real-time) and conceptual simplicity of the method, the
results are very good.
The middle panels of Fig. 8.15 show the error rates for the method of Kla-
puri [354]. The first detected FO was used for the predominant FO estimation.
In these experiments, the pitch range was limited to five octaves between
65 Hz and 2.1 kHz. The method performs robustly in all cases and is very
accurate, especially in the 93 ms analysis frame. Computational complexity is
a drawback of this method. The calculations are clearly slower than real-time
on a 2-GHz desktop computer, the most intensive part being the cochlear
filterbank and the within-band DFT calculations.
The right-hand panels of Fig. 8.15 show the error rates for a state-of-the-
art reference method proposed by Klapuri in [351]. This method is based on
262 Anssi Klapuri

Tolonen et al. Tolonen et al. Klapuri

Polyphony 1 Polyphony 4 Polyphony 1

ll
0 20 40 60 80 20 40 60 80
EM
0 20 40 60 80 20 40 60 80
Time interval (ms) Time interval (ms) Time interval (ms) Time interval (ms)

Fig. 8.16. Error rates as a function of the interval between the sound onset and
the beginning of a 46 ms analysis frame. The two panels on the left show results
for the method of Tolonen and Karjalainen [627] and the two panels on the right
for the method of Klapuri [354]. The black bars show multiple FO estimation error
rates and the white bars show predominant FO estimation error rates.

spectral techniques instead of an auditory model and is therefore a good point

of comparison. T h e test cases given to Klapuri's method [354] and the refer-
ence m e t h o d were identical. It was observed t h a t the reference m e t h o d requires
quite a long analysis frame to resolve and process the overtones of low-pitched
sounds, and mallet percussion instruments could not be reliably analysed. In
addition to the differences in handling t h e higher-order overtones, a factor
involved is t h a t the frequency resolution of the Fourier spectrum is linear,
whereas time-domain periodicity analysis within the auditory channels leads
to 1 / / frequency resolution, which enables more accurate analysis at the lower
end of the logarithmic scales applied in music. T h e reference m e t h o d is con-
ceptually (technically) the most complex among the three.
An important factor in the above results is t h a t the analysis frames were
positioned immediately after t h e onsets of t h e sounds. Figure 8.16 shows the
error rates of the two methods as a function of the time interval between
the sound onset and the beginning of the analysis frame. As can be seen,
the error rates improve clearly as the interval increases, and especially the
predominant FO estimation error rates shrink to about a third of the initial
values after 80 ms of the onset. This is because the noisy beginning transients
of many sounds die off rapidly and FO estimation becomes easier thereafter.
In music signals, however, notes are often short and such an offset cannot
be applied. In Fig. 8.15, maximally realistic simulations were of interest and
thus a zero offset was applied. Figure 8.16 shows results only for the 46 ms
analysis frame, b u t the general trend is similar (although less pronounced) for
the longer frame.

8 . 5 . 5 S u m m a r y of t h e M u l t i p l e FO E s t i m a t i o n M e t h o d s

T h e beginning of this section listed several issues where the pitch perception
models fall short of being practically applicable multiple FO estimators. This
section summarizes and discusses the various technical solutions t h a t were
proposed as improvements.
 8 Auditory Model-Based Methods for Multiple FO Estimation 263

Two main approaches can be distinguished among the techniques used to

extend a single-pitch model to the estimation of multiple pitches: the iterative
estimation-and-cancellation approach and the joint estimation approach. Most
methods fall into the former category: FO estimation is done using the sum-
mary ACF, for example, and the FO found is then cancelled before deciding
the next one. Meddis and Hewitt performed the cancellation by removing the
auditory channels associated with the first detected pitch [458]. De Cheveigne
employed within-channel cancellation filtering in the time domain [130]. Kla-
puri subtracted the partials of a detected sound in the frequency domain and
removed only the lower-order partials entirely [354].
Joint estimation methods were proposed by de Cheveigne [129], Karjalainen
and Tolonen [327], and Wu et al. [677]. Among these, the method of de
Cheveigne was not actually based on an auditory model, but the method
applied two cancellation filters in a cascade and searched for such cancellation-
filter periods that the output power was minimized. Karjalainen and Tolonen
enhanced the summary ACF so that all FOs could be directly extracted from
the result. In the method of Wu et al., the distribution of the peaks in the
sub-band ACFs was statistically modelled in the cases of zero, one, or two
pitches.
The limited robustness of the pitch perception models in polyphonic sig-
nals is another important problem addressed by the multiple FO estimation
methods. The chord-root detection problem was mentioned as an example
of this. The SACF enhancing technique of Karjalainen and Tolonen [627] is
rather efficient in this respect, as illustrated in Fig. 8.11. Klapuri addressed
the problem by applying the lowpass response in (8.25), which suppresses the
support of higher-order partials to the chord root unless the series of partials
has sufficiently uniform amplitudes so as to generate strong beating at the
fundamental rate. This is usually not the case if the partials are due to sev-
eral different sounds (component FOs of a chord). Also, the use of harmonic
selection in the frequency domain alleviated the interference of other sounds
since the spectrum between the partials was not used in salience calculations.
Iterative estimation and cancellation methods that estimate the first FO di-
rectly from the summary ACF suffer from its robustness limitations [458],
[130].
Robustness for different sound sources (different musical instruments) is a
very important aspect in FO estimation. Here the pitch perception models are
readily very efficient. Meddis's hair-cell model compresses the sub-band sig-
nals and results in spectral whitening, that is, removal of timbral information
to some extent [457], [456]. Ellis [171] and Klapuri [354] carried out this func-
tion by scaling the sub-band signals inversely proportional to their variance.
Karjalainen and Tolonen pre-processed the input signals by inverse warped-
hnear-prediction filtering. This had the advantage that a multi-channel fil-
terbank was not needed [327], [272]. An advantage of all these is that they
fiatten the spectral energy distribution without raising the noise floor in re-
lation to spectral peaks. The latter happens for example in cepstrum pitch
264 Anssi Klapuri

detection, where the logarithm function is apphed bin-by-bin to the magni-

tude spectrum [535]. The system of Wu et al. [677] is interesting, since in this
method, whitening would not have any effect at all because only the lag-values
of within-channel ACF peaks are retained and not their amplitudes.
Noise robustness was not discussed in depth in this chapter. In music, per-
cussive instruments and recording imperfections cause noise-like interference
for the FO estimation. Particular emphasis on this issue was laid by Wu et al.,
who performed channel and peak selection so as to avoid the spectro-temporal
regions that were severely corrupted by noise. The authors remarked that they
essentially treated multiple FO tracking and noise robustness as a single prob-
lem [677, p. 240]. Karjalainen and Tolonen selected the generalized ACF power
so as to make a compromise between noise robustness and spectral flattening
[327].
Computational complexity of the pitch models was significantly reduced
only in the method of Karjalainen and Tolonen [327]. In the other methods,
the most time-consuming operation is typically the peripheral filterbank and
the periodicity analysis within channels, usually leading to computation times
which are 10 to 100 times slower than that of the method of Karjalainen and
Tolonen. Ellis computed the within-channel ACFs only for a set of logarith-
mically distributed lag values, which allowed the use of a very good time
resolution without causing a prohibit at ive computational load [171]. In the
iterative methods, the peripheral analysis usually has to be computed only
once [458], [354].

8.6 Conclusions

Pitch perception models and practical FO estimators address sHghtly different

tasks and are judged according to different criteria. The former should faith-
fully represent the mechanisms of the human auditory system, whereas the
latter are expected to perform accurate multiple FO estimation by any means
available. The main focus of this chapter was on the practical side. However,
the two aspects have significantly influenced and beneflted each other and this
is one reason to study auditory modelling.
Many characteristics of human pitch perception can be traced to the pe-
ripheral stages of hearing, as discussed in Section 8.3. In this sense, auditory
models have a lot to say about the intermediate data representations used in
acoustic signal analysis. A particularly important principle in an auditorily
motivated analysis is that the higher-order overtones of a sound are processed
collectively within each auditory channel; estimation and separation of indi-
vidual higher-order partials is not attempted. The cochlear filterbank is 'fair'
for different FO values in this respect since the first few harmonic partials of
all FOs are resolved into separate auditory channels, whereas the harmonics
above about 10 go to the same channel along with their neighbours and gen-
erate amplitude envelope beating at the fundamental rate. This is an efficient
 8 Auditory Model-Based Methods for Multiple FO Estimation 265

mechanism for dealing with the higher-order overtones in complex polyphonic

data. The other advantages and disadvantages of a perceptually motivated
data representation were discussed in Section 8.4.4.
Compared with the peripheral stages of hearing, at least an equally im-
portant part of pitch perception takes place in the brain (Steps 3 and 4 in the
overview on p. 235). These stages are not yet well understood and thus there
is a larger variance in the proposed practical techiques as well. The biggest
defect of the existing pitch perception models from the music transcription
viewpoint is that they have been designed to process isolated sounds instead of
polyphonic signals. Different techniques for transforming a pitch model into a
multiple FO estimator were described in Section 8.5. Both iterative methods
and joint estimation methods were discussed, and different ways of cancelling
a detected FO from the mixture signal were described. Periodicity analysis
techniques were presented that applied the ACF [458], [130], the generalized
ACF [627], statistical modelling of the ACF peaks [677], adaptive oscillators
[434], or simulation of comb filters in the frequency domain [354]. For now,
none of the described methods can claim to be the 'right' or the optimal one,
but they provide a wealth of technical solutions and approaches to build upon.
9
Unsupervised Learning Methods for Source
Separation in Monaural Music Signals

Tuomas Virtanen

Institute of Signal Processing, Tampere University of Technology,

Korkeakoulunkatu 1, 33720 Tampere, Finland
Tuomas.VirtanenQtut.fi

9.1 Introduction
Computational analysis of polyphonic musical audio is a challenging problem.
When several instruments are played simultaneously, their acoustic signals
mix, and estimation of an individual instrument is disturbed by the other co-
occurring sounds. The analysis task would become much easier if there was
a way to separate the signals of different instruments from each other. Tech-
niques that implement this are said to perform sound source separation. The
separation would not be needed if a multi-track studio recording was available
where the signal of each instrument is on its own channel. Also, recordings
done with microphone arrays would allow more efficient separation based on
the spatial location of each source. However, multi-channel recordings are usu-
ally not available; rather, music is distributed in stereo format. This chapter
discusses sound source separation in monaural music signals, a term which
refers to a one-channel signal obtained by recording with a single microphone
or by mixing down several channels.
There are many signal processing tasks where sound source separation
could be utilized, but the performance of the existing algorithms is still quite
limited compared to the human auditory system, for example. Human listeners
are able to perceive individual sources in complex mixtures with ease, and
several separation algorithms have been proposed that are based on modelling
the source segregation ability in humans (see Chapter 10 in this volume).
Recently, the separation problem has been addressed from a completely
different point of view. The term unsupervised learning is used here to char-
acterize algorithms which try to separate and learn the structure of sources
in mixed data based on information-theoretical principles, such as statisti-
cal independence between sources, instead of sophisticated modelling of the
source characteristics or human auditory perception. Algorithms discussed in
this chapter are independent component analysis (ICA), sparse coding, and
non-negative matrix factorization (NMF), which have been recently used in
268 Tuomas Virtanen

source separation tasks in several application areas. When used for monau-
ral audio source separation, these algorithms usually factor the spectrogram
or other short-time representation of the input signal into elementary com-
ponents, which are then clustered into sound sources and further analysed to
obtain musically important information. Although the motivation of unsuper-
vised learning algorithms is not in the human auditory perception, there are
similarities between them. For example, all the unsupervised learning meth-
ods discussed here are based on reducing redundancy in data, and it has been
found that redundancy reduction takes place in the auditory pathway, too [85].
The focus of this chapter is on unsupervised learning algorithms which
have proven to produce applicable separation results in the case of music
signals. There are some other machine learning algorithms which aim at sep-
arating speech signals based on pattern recognition techniques, for example
[554].
All the algorithms mentioned above (ICA, sparse coding, and NMF) can
be formulated using a linear signal model which is explained in Section 9.2.
Different data representations are discussed in Section 9.2.2. The estimation
criteria and algorithms are discussed in Sections 9.3, 9.4, and 9.5. Methods
for obtaining and utilizing prior information are presented in Section 9.6.
Once the spectrogram is factored into components, these can be clustered
into sound sources or further analysed to obtain musical information. The
post-processing methods are discussed in Section 9.7. Systems extended from
the linear model are discussed in Section 9.8.

9.2 Signal Model

When several sound sources are present simultaneously, the acoustic wave-
forms of the individual sources add linearly. Sound source separation is defined
as the task of recovering each source signal from the acoustic mixture. A com-
plication is that there is no unique definition for a sound source. One possi-
bility is to consider each vibrating physical entity, for example each musical
instrument, as a sound source. Another option is to define this according to
what humans tend to perceive as a single source. For example, if a violin, sec-
tion plays in unison, the violins are perceived as a single source, and usually
there is no need to separate the signals played by each violin. In Chapter 10,
these two alternatives are referred to as physical source and perceptual source,
respectively (see p. 302). Here we do not specifically commit ourselves to either
of these. The type of the separated sources is determined by the properties
of the algorithm used, and this can be partly affected by the designer accord-
ing to the application at hand. In music transcription, for example, all the
equal-pitched notes of an instrument can be considered as a single source.
Many unsupervised learning algorithms, for example standard ICA, require
that the number of sensors be larger or equal to the number of sources. In
multi-channel sound separation, this means that there should be at least as
 9 Unsupervised Learning Methods for Source Separation 269

many microphones as there are sources. However, automatic transcription of

music usually aims at finding the notes in monaural (or stereo) signals, for
which basic ICA methods cannot be used directly. By using a suitable signal
representation, the methods become applicable with one-channel data.
The most common representation of monaural signals is based on short-
time signal processing, in which the input signal is divided into (possibly over-
lapping) frames. Frame sizes between 20 and 100 ms are typical in systems
designed to separate musical signals. Some systems operate directly on time-
domain signals and some others take a frequency transform, for example
the discrete Fourier transform (DFT) of each frame. The theory and general
discussion of time-frequency representations is presented in Chapter 2.

9.2.1 Basis Functions and Gains

The representation of the input signal within each frame ^ = 1 . . . T is denoted

by an observation vector Xt. The methods presented in this chapter model Xt
as a weighted sum of basis functions b^, TI = 1 . . . AT, so that the signal model
can be written as
N
yit^^gn,thn, t-l,...,r, (9.1)
n=l

where N <^ T is the number of basis functions, and gn,t is the amount of
contribution, or gain, of the n*^ basis function in the t*^ frame. Some methods
estimate both the basis functions and the time-varying gains from a mixed
input signal, whereas others use pre-trained basis functions or some prior
information about the gains.
The term component refers to one basis function together with its time-
varying gain. Each sound source is modelled as a sum of one or more compo-
nents, so that the model for source m in frame t is written as

ym,t = Yl 9n,thn^ (9.2)

neSm

where Sm is the set of components within source m. The sets are disjoint, i.e.,
each component belongs to only one source.
In (9.1) approximation is used, since the model is not necessarily noise-free.
The model can also be written with a residual term r^ as
N
^t = Yl ^^'*^^ + ^*' t = 1,..., T. (9.3)
n=l

By assuming some probability distribution for the residual and a prior distri-
bution for other parameters, a probabilistic framework for the estimation of
hn and gn,t can be formulated (see e.g. Section 9.4). Here (9.1) without the
270 Tuomas Virtanen

residual term is preferred for its simplicity. For T frames, the model (9.1) can
be written in matrix form as
X BG, (9.4)
where X = [ x i , X 2 , . . . , X T ] is the observation matrix^ B = [bi, b 2 , . . . , bjv]
is the mixing matrix, and [G]n,t 9n,t is the gain matrix. The notation [G]n,t
is used to denote the (n, t)*^ entry of matrix G. The term mixing matrix is
typically used in ICA, and here we follow this convention.
The estimation algorithms can be used with several data representations.
Often the absolute values of the DFT are used; this is referred to as the mag-
nitude spectrum in the following. In this case, x^ is the magnitude spectrum
within frame t, and each component n has a fixed magnitude spectrum b^
with a time-varying gain gn^t- The observation matrix consisting of framewise
magnitude spectra is here called a magnitude spectrogram. Other representa-
tions are discussed in Section 9.2.2.
The model (9.1) is flexible in the sense that it is suitable for represent-
ing both harmonic and percussive sounds. It has been successfully used in
the transcription of drum patterns [188], [505] (see Chapter 5), in the pitch
estimation of speech signals [579], and in the analysis of polyphonic music
signals [73], [600], [403], [650], [634], [648], [43], [5].
Figure 9.1 shows an example signal which consists of a diatonic scale and
a C major chord played by an acoustic guitar. The signal was separated into
components using the NMF algorithm described in [600], and the resulting
components are depicted in Fig. 9.2. Each component corresponds roughly to
one fundamental frequency: the basis functions are approximately harmonic
and the time-varying gains follow the amplitude envelopes of the notes. The
separation is not perfect because of estimation inaccuracies. For example, in
some cases the gain of a decaying note drops to zero when a new note begins.
Factorization of the spectrogram into components with a fixed spectrum
and a time-varying gain has been adopted as a part of the MPEG-7 pattern
recognition framework [72], where the basis functions and the gains are used
as features for classification. Kim et al. [341] compared these to mel-frequency
cepstral coefficients which are commonly used features in the classification of
audio signals. In this study, mel-frequency cepstral coefficients performed bet-
ter in the recognition of sound effects and speech than features based on ICA
or NMF. However, final conclusions about the apphcability of these methods
to sound source recognition have yet to be made. The spectral basis decompo-
sition specified in MPEG-7 models the summation of components on a decibel
scale, which makes it unlikely that the separated components correspond to
physical sound objects.

9.2.2 Data Representation

The model (9.1) presented in the previous section can be used with time-
domain or frequency-domain observations and basis functions. Time-domain
 9 Unsupervised Learning Methods for Source Separation 271

5000

4500

4000

3500

3000

o 2500

"2000

1500

1000

500

0
1 1.5 2 3.5
Time (seconds)

Fig. 9.1. Spectrogram of an example signal which consist of a diatonic scale from
C5 to C6, followed by a C major chord (simultaneous notes C5, E4, and G5), played
by an acoustic guitar. The notes are not damped, meaning that consecutive notes
overlap.

observation vector Xt is the signal within frame t directly, whereas a frequency-

domain observation vector is obtained by applying a chosen transformation to
this. The representation of the signal and the basis functions have to be the
same. ICA and sparse coding allow the use of any short-time signal represen-
tation, whereas for NMF, only a frequency-domain representation is appro-
priate. Naturally, the representation has a significant effect on performance.
The advantages and disadvantages of different representations are considered
in this section. For a more extensive discussion, see Casey [70] or Smaragdis
[598].

Time-Domain Representation

Time-domain representations are straightforward to compute, and all the in-

formation is preserved when an input signal is segmented into frames and win-
dowed. However, time-domain basis functions are problematic in the sense that
a single basis function alone cannot represent a meaningful sound source: the
phase of the signal within each frame varies depending on the frame position.
In the case of a short-duration percussive source, for example, a separate basis
function is needed for every possible position of the sound event within the
272 Tuomas Virtanen

V^A,.... ^^^^ AAAAAMAAAA J

r\MA(WM^A./\^ /LA

M\A,KAAAAAA

ffiO^M^
N\f^^^^r\A^/v.^^^^ni
AJA/JW/\^JKKJ\J\

KflAA/vAAAvAXJ
^/lA.AA/wwu^.
i ^ ^ t U fli i. lA/irYTYiAnriAnr^^-^^^^
1 2 3 1000 2000 3000 4000
Time (seconds) Frequency (Hertz)

Fig. 9.2. Components estimated from the example signal in Fig. 9.1. Basis functions
are plotted on the right and the corresponding time-varying gains on the left. Each
component except the bottom one corresponds to an individual pitch value and the
gains follow roughly the amplitude envelope of each note. The bottom component
models the attack transients of the notes. The components were estimated using the
NMF algorithm [400], [600] and the divergence objective (explained in Section 9.5).

frame. A shift-invariant model which is later discussed in Section 9.8 is one

possible method of overcoming this limitation [43].
The time-domain signals of real-world sound sources are generally not
identical at different occurrences since the phases behave very irregularly. For
example, the overtones of a pitched musical instrument are not necessarily
phase-locked, so that the time-domain waveform varies over time. Therefore,
one has to use multiple components to represent even a single note of a pitched
instrument. In the case of percussive sound sources, this phenomenon is even
clearer: the time-domain waveforms vary a lot at different occurrences.
The larger the number of the components, the more uncertain is their
estimation and further analysis, and the more observations are needed. If the
sound event represented by a component occurs only once in the input signal,
separating it from co-occurring sources is difficult since there is no information
about the component elsewhere in the signal. Also, clustering the components
into sources becomes more difficult when there are many of them for each
source.
 9 Unsupervised Learning Methods for Source Separation 273

Separation algorithms which operate on time-domain signals have been

proposed for example by Dubnov [157], Jang and Lee [314], and Blumensath
and Davies [43]. Abdallah and Plumbley [3], [2] found that the independent
components analysed from time-domain music and speech signals were similar
to a wavelet or short-time DFT basis. They trained the basis functions using
several days of radio output from BBC Radio 3 and 4 stations.

Frequency-Domain Representation

When using a frequency transform such as the DFT, the phases of the
complex-valued transform can be discarded by considering only the mag-
nitude or power spectrum. Even though some information is lost, this also
eliminates the phase-related problems of time-domain representations. Unlike
time-domain basis functions, many real-world sounds can be rather well ap-
proximated with a fixed magnitude spectrum and a time-varying gain, as seen
in Figs. 9.1 and 9.2, for example. Sustained instruments in particular tend to
have a stationary spectrum after the attack transient.
In most systems aimed at the separation of sound sources, DFT and a
fixed window size is applied, but the estimation algorithms allow the use
of any time-frequency representation. For example, a logarithmic spacing of
frequency bins has been used [58], which is perceptually and musically more
plausible than a constant spectral resolution.
The linear summation of time-domain signals does not imply the linear
summation of their magnitude or power spectra, since phases of the source
signals affect the result. When two signals sum in the time domain, their
complex-valued DFTs sum Hnearly, X{k) = Yi{k) + Y2{k), but this equality
does not apply for the magnitude or power spectra. However, provided that
the phases of Yi{k) and Y2{k) are uniformly distributed and independent of
each other, we can write

E{\Xik)f} = \Y^ik)\' + \Y2{k)\\ (9.5)

where E{'} denotes expectation. This means that in the expectation sense,
we can approximate time-domain summation in the power spectral domain, a
result which holds for more than two sources as well. Even though magnitude
spectrogram representation has been widely used and it often produces good
results, it does not have similar theoretical justification. Since the summation
is not exact, use of phaseless basis functions causes an additional source of
error. Also, a phase generation method has to be implemented if the sources
are to be synthesized separately. These are discussed in Section 9.7.3.
The human auditory system has a large dynamic range: the difference
between the threshold of hearing and the threshold of pain is approximately
100 dB [550]. Unsupervised learning algorithms tend to be more sensitive to
high-energy observations. If sources are estimated from the power spectrum,
some methods fail to separate low-energy sources even though they would be
274 Tuomas Virtanen

perceptually and musically meaningful. This problem has been noticed, e.g.,
by FitzGerald in the case of percussive source separation [186, pp. 93-100].
To overcome the problem, he used an algorithm which processed separately
high-frequency bands which contain low-energy sources, such as hi-hats and
cymbals [187]. Vincent and Rodet [648] addressed the same problem. They
proposed a model in which the noise was additive in the log-spectral domain.
The numerical range of a logarithmic spectrum is compressed, which increases
the sensitivity to low-energy sources. Additive noise in the log-spectral domain
corresponds to multiplicative noise in power spectral domain, which was also
assumed in the system proposed by Abdallah and Plumbley [5]. Virtanen
proposed the use of perceptually motivated weights [651]. He used a weighted
cost function in which the observations were weighted so that the quantitative
significance of the signal within each critical band was equal to its contribution
to the total loudness.

9.3 Independent Component Analysis

ICA has been successfully used in several 'blind' source separation tasks, where
very httle or no prior information is available about the source signals. One
of its original target applications was multi-channel sound source separation,
but it has also had several other uses. ICA attempts to separate sources by
identifying latent signals that are maximally independent. In practice, this
usually leads to the separation of meaningful sound sources.
Mathematically, statistical independence is defined in terms of probabil-
ity densities: random variables x and y are said to be independent if their
joint probability distribution function^ ?{x^y) is a product of the marginal
distribution functions, p{x,y) = p{x)p{y).
The dependence between two variables can be measured in several ways.
Mutual information is a measure of the information that given random vari-
ables have on some other random variables [304]. The dependence is also
closely related to the Gaussianity of the distribution of the variables. Accord-
ing to the central limit theorem, the distribution of the sum of independent
variables is more Gaussian than their original distributions, under certain con-
ditions. Therefore, some ICA algorithms aim at separating output variables
whose distributions are as far from Gaussian as possible.
The signal model in ICA is linear: K observed variables x i , . . . , XK are
modelled as linear combinations of A^ source variables ^ i , . . . , ^ A T - In a vector-
matrix form, this can be written as

X - Bg, (9.6)

where x = [XI,...XK] is an observation vector, [B]fc,n = ^fc,n is a mixing

matrix, and g = [^ri,..., g^] is a source vector. Both B and g are unknown.
^The concept of probability distribution function is described in Chapter 2.
 9 Unsupervised Learning Methods for Source Separation 275

The standard ICA requires that the number of observed variables K (the
number of sensors) be equal to the number of sources A^. In practice, the num-
ber of sensors can also be larger than the number of sources, because the vari-
ables are typically decorrelated using principal component analysis (PCA; see
Chapter 2), and if the desired number of sources is less than the number of
variables, only the principal components corresponding to the largest eigen-
values are selected.
As another pre-processing step, the observed variables are usually centred
by subtracting the mean and their variance is normalized to the unity. The
centred and whitened data observation vector x is obtained from the original
observation vector x by
x = V(x-/i), (9.7)
where fj, is the empirical mean of the observation vector, and V is a whitening
matrix, which is often obtained from the eigenvalue decomposition of the
empirical covariance matrix of the observations [304]. The empirical mean
and covariance matrix are explained in Chapter 2.
To simplify the notation, it is assumed that the data x in (9.6) is already
centred and decorrelated, so that K = N. The core ICA algorithm carries
out the estimation of an unmixing matrix W ~ B~^, assuming that B is
invertible. Independent components are obtained by multiplying the whitened
observations by the estimate of the unmixing matrix, to result in the source
vector estimate g:
g = Wx. (9.8)
The matrix W is estimated so that the output variables, i.e., the elements
of g, become maximally independent. There are several criteria and algo-
rithms for achieving this. The criteria, such as non-Gaussianity and mutual
information, are usually measured using high-order cumulants such as kurto-
sis, or expectations of other non-quadratic functions [304]. ICA can be also
viewed as an extension of PCA. The basic PCA decorrelates variables so that
they are independent up to second-order statistics. It can be shown that if
the variables are uncorrelated after taking a suitable non-linear function, the
higher-order statistics of the original variables are independent, too. Thus,
ICA can be viewed as a non-linear decorrelation method.
Compared with the previously presented linear model (9.3), the standard
ICA model (9.6) is exact, i.e., it does not contain the residual term. Some
special techniques can be used in the case of the noisy signal model (9.3),
but often noise is just considered as an additional source variable. Because of
the dimension reduction with PCA, E g gives an exact model for the PCA-
transformed observations but not necessarily for the original ones.
There are several ICA algorithms, and some implementations are freely
available, such as FastICA [302], [182] and JADE [65]. Computationally quite
efficient separation algorithms can be implemented based on FastICA, for
example.
276 Tuomas Virtanen

9.3.1 Independent Subspace Analysis

The idea of independent subspace analysis (ISA) was originally proposed by

Hyvarinen and Hoyer [303]. It combines the multidimensional ICA with in-
variant feature extraction, which are shortly explained later in this section.
After the work of Casey and Westner [73], the term ISA has been commonly
used to denote techniques which apply ICA to factor the spectrogram of a
monaural audio signal to separate sound sources. ISA provides a theoretical
framework for the whole separation procedure described in this chapter, in-
cluding spectrogram representation, decomposition by ICA, and clustering.
Some authors use the term ISA also to refer to methods where some other
algorithm than ICA is used for the factorization [648].
The general ISA procedure consists of the following steps:

1. Calculate the magnitude spectrogram X (or some other representation)

of the input signal.
2. Apply PCA^ on the matrix X of size {K x T) to estimate the number
of components N and to obtain whitening and dewhitening matrices V
and V"^, respectively. A centred, decorrelated, and dimensionally reduced
observation matrix X of size {N X T) is obtained as X = V(X - /xl"^),
where 1 is a all-ones vector of length T.
3. Apply ICA to estimate an unmixing matrix W . B and G are obtained as
B = W - i and G = W X .
4. Invert the decorrelation operation in Step 2 in order to get the mixing
matrix B = V"^B and source matrix G G + W V / L A 1 ^ for the original
observations X.
5. Cluster the projected components to sources (see Section 9.7.1).

The above steps are explained in more detail below. Depending on the appli-
cation, not all of them may be necessary. For example, prior information can
be used to set the number of components in Step 2.
The basic ICA is not directly suitable for the separation of one-channel
signals, since the number of sensors has to be larger than or equal to the
number of sources. Short-time signal processing can be used in an attempt
to overcome this limitation. Taking a frequency transform such as DFT, each
frequency bin can be considered as a sensor which produces an observation in
each frame. With the standard linear ICA model (9.6), the signal is modelled
as a sum of components, each of which has a static spectrum (or some other
basis function) and a time-varying gain.
The spectrogram factorization has its motivation in invariant feature ex-
traction, which is a technique proposed by Kohonen [356]. The short-time
spectrum can be viewed as a set of features calculated from the input signal.
As discussed in Section 9.2.2, it is often desirable to have shift-invariant basis
^Singular value decomposition can also be used to estimate the number of com-
ponents [73].
 9 Unsupervised Learning Methods for Source Separation 277

functions, such as the magnitude or power spectrum [356], [303]. Multidimen-

sional ICA (explained below) is used to separate phase-invariant features into
invariant feature subspaces, where each source is modelled as the sum of one
or more components [303].
Multidimensional ICA [64] is based on the same linear generative model
(9.6) as ICA, but the components are not assumed to be mutually indepen-
dent. Instead, it is assumed that the components can be divided into disjoint
sets, so that the components within each set may be dependent on each other,
while dependencies between sets are not allowed. One approach to estimat-
ing multidimensional independent components is to first apply standard ICA
to estimate the components, and then group them into sets by measuring
dependencies between them.'^
ICA algorithms aim at maximizing the independence of the elements of the
source vector g = W x . In ISA, the elements correspond to the time-varying
gains of each component. However, the objective can also be the independence
of the spectra of components, since the roles of the mixing matrix and gain
matrix can be swapped by X = B G <^ X^ = G ^ B ^ . The independence
of both the time-varying gains and basis functions can be obtained by using
the spatiotemporal ICA algorithm [612]. There are no exhaustive studies re-
garding different independence criteria in monaural audio source separation.
Smaragdis argued that in the separation of complex sources, the criterion of
independent time-varying gains is better, because of the absence of consis-
tent spectral characteristics [598]. FitzGerald reported that the spatiotempo-
ral ICA did not produce significantly better results than normal ICA, which
assumes the independence of gains or spectra [186].
The number of frequency channels is usually larger than the number of
components to be estimated with ICA. PC A or singular value decomposition
(SVD) of the spectrogram can be used to estimate the number of components
automatically. SVD decomposes the spectrogram into a sum of components
with a fixed spectrum and time-varying gain, so that the spectra and gains of
different components are orthogonal, whereas PCA results in the orthogonality
of either the spectra or the gains. The components with the largest singular
values are chosen so that the sum of their singular values is larger than or
equal to a pre-defined threshold 0 < ^ < 1 [73].
ISA has been used for general audio separation by Casey and Westner [73],
for the analysis of musical trills by Brown and Smaragdis [58], and for per-
cussion transcription by FitzGerald et al. [187], to mention some examples.

9.3.2 Non-Negativity Restrictions

When magnitude or power spectrograms are used, the basis functions are
magnitude or power spectra which are non-negative by definition. Therefore,

^ICA aims at maximizing the independence of the output variables, but it cannot
guarantee their complete independence, as this depends also on the input signal.
278 Tuomas Virtanen

it can be advantageous to restrict the basis functions to be entry-wise non-

negative. Also, it may be useful not to allow negative gains, but to constrain
the components to be purely additive. Standard ICA is problematic in the
sense that it does not enable these constraints. In practice, ICA algorithms
also produce negative values for the basis functions and gains, and often there
is no physical interpretation for such components.
ICA with non-negativity restrictions has been studied for example by
Plumbley and Oja [526], and the topic is currently under active research.
Existing non-negative ICA algorithms can enforce non-negativity for the gain
matrix but not for the mixing matrix. They also assume that the probability
distribution of the source variables gn is non-zero all the way down to zero,
i.e., the probability Qn < S is non-zero for any ^ > 0. The algorithms are based
on a noise-free mixing model and in our experiments with audio spectrograms,
they tended to be rather sensitive to noise.
It has turned out that the non-negativity restrictions alone are sufficient
for the separation of the sources, without the expHcit assumption of statistical
independence. NMF algorithms are discussed in Section 9.5.

9.4 Sparse Coding

Sparse coding represents a mixture signal in terms of a small number of active

elements chosen out of a larger set [486]. This is an efficient approach for learn-
ing structures and separating sources from mixed data. General discussion of
sparse adaptive representations suitable for the analysis of musical signals is
given in Chapter 3. In the hnear signal model (9.4), the sparseness restriction
is usually applied on the gains G, which means that the probability of an
element of G being zero is high. As a result, only a few components are active
at a time and each component is active only in a small number of frames. In
musical signals, a component can represent, e.g., all the equal-pitched notes
of an instrument. It is likely that only a small number of pitches are played
simultaneously, so that the physical system behind the observations generates
sparse components.
In this section, a probabilistic framework is presented, where the source
and mixing matrices are estimated by maximizing their posterior distribu-
tions. The framework is similar with the one presented by Olshausen and
Field [486]. Several assumptions of, e.g., the noise distribution and prior dis-
tribution of the gains are used. Obviously, different results are obtained by
using different distributions, but the basic idea is the same. The method pre-
sented here is also closely related to the algorithms proposed by Abdallah
and Plumbley [4] and Virtanen [650], which were used in the analysis of music
signals.
The posterior distribution of B and G given an observed spectrogram X
is denoted by p(B, G|X). The maximization of this can be formulated as [339,
p. 351]
 9 Unsupervised Learning Methods for Source Separation 279

maxp(B, G|X) oc maxp(X|B, G)p(B, G), (9.9)

B,G B,G
where p(X|B,G) is the probabihty of observing X given B and G, and
p(B,G) is the joint prior distribution of B and G. The concepts of proba-
bihty distribution function, conditional probabihty distribution function, and
maximum a posteriori estimation are described in Chapter 2.
For mathematical tractability, it is typically assumed that the noise (the
residual term in (9.3)) is i.i.d.; independent from the model B G , and normally
distributed with variance cr^ and zero mean. The likelihood of B and G (see
Section 2.2.5 for the eplanation of likelihood functions) can be written as

p(X|B,G) = n ^ - p (^_([X]M-JBG].,)^-| (9 10)

It is further assumed here that B has a uniform prior, so that p(B, G) (x

p(G). Each time-varying gain [G]n,t is assumed to have a sparse probability
distribution function of the exponential form

P([Gkt) = | e x p ( - / ( [ G ] , 0 ) - (9.11)

A normalization factor Z has to be used so that the density function sums to

unity. The function / is used to control the shape of the distribution and is
chosen so that the distribution is uni-modal and peaked at zero with heavy
tails. Some examples are given later.
For simplicity, all the entries of G are assumed to be independent from
each other, so that the probability distribution function of G can be written
as a product of the marginal densities:

P(G) = n i e x p ( - / ( [ G W ) ) . (9.12)
n.t

It is obvious that in practice the gains are not independent of each other,
but this approximation is done to simplify the calculations. From the above
definitions we get

{[X]k,t - [BG]k,t)'
max p(B, G|X) ex max TT - = exp (
B,<^ B,Lr -^f- ay Z7T \ 2<T2

(9.13)
t,k

By taking a logarithm, then,tproducts become summations, and the exp-

operators and scaling terms can be discarded. This can be done since logarithm
is order preserving and therefore does not affect the maximization. The sign
is changed to obtain a minimization problem
280 Tuomas Virtanen

mm E ^""^'"^ ~if''^''^' + E /([G]n.t), (9.14)

B,G t,k n,t

which can be written as

(9.15)
n,t

where the Frobenius norm of a matrix is defined as

\nF=l^[Y]i (9.16)
^,J

In (9.15), the function / is used to penahze 'active' (non-zero) entries of

G. For example, Olshausen and Field [486] suggested the functions f{x) =
log(l -f x^), f{x) = |x|, and f{x) = x^. In audio source separation, Benaroya
et al. [32] and Virtanen [650] have used f{x) \x\. The prior distribution
used by Abdallah and Plumbley [2], [4] corresponds to the function

\x\ > /i,

f{x) = m^ (9.17)
I/i(l a ) - f Q;|X|, |X| < / i .

where the parameters /i and a control the relative mass of the central peak in
the prior, and the term fi{l a) is used to make the function continuous at x =
/i. All these functions give a smaller cost and a higher prior probability for
gains near zero. The cost function f{x) = \x\ and the corresponding Laplacian
prior p(x) = | e x p ( - | x | ) are illustrated in Fig. 9.3. Systematic large-scale
evaluations of different sparse priors in audio signals have not been carried
out. Naturally, the distributions depend on source signals, and also on the
data representation.

Fig. 9.3. The cost function f{x) = \x\ (left) and the corresponding Laplacian prior
distribution p{x) = |exp(|a::|) (right). Values of G near zero are given a smaller
cost and a higher probability.
 9 Unsupervised Learning Methods for Source Separation 281

Prom (9.15) and the above definitions of / , it can be seen that a sparse
representation is obtained by minimizing a cost function which is the weighted
sum of the reconstruction error term ||X BG||fr and the term which incurs
a penalty on non-zero elements of G. The variance a^ is used to balance
between these two. This objective 9.15 can be viewed as a penalized likelihood,
discussed in the Tools section (see Sections 2.2.9 and 2.3.3).
Typically, / increases monotonically as a function of the absolute value of
its argument. The presented objective requires that the scale of either the basis
functions or the gains is somehow fixed. Otherwise, the second term in (9.15)
could be minimized without affecting the first term by setting B ^ B ^ and
G ^ G/^, where the scalar 0 -^ oo. The scale of the basis functions can
be fixed for example with an additional constraint ||bn|| = 1, as done by
Hoyer [299], or the variance of the gains can be fixed.
The minimization problem (9.15) is usually solved using iterative algo-
rithms. If both B and G are unknown, the cost function may have several local
minima, and in practice reaching the global optimum in a limited time cannot
be guaranteed. Standard optimization techniques based on steepest descent,
covariant gradient, quasi-Newton, and active-set methods can be used. Differ-
ent algorithms and objectives are discussed for example by Kreutz-Delgado
et al. [373].
If B is fixed, more efficient optimization algorithms can be used. This
can be the case for example when B is learned in advance from training
material where sounds are presented in isolation. These methods are discussed
in Section 9.6.
No methods have been proposed for estimating the number of sparse com-
ponents in a monaural audio signal. Therefore, N has to be set either manu-
ally, using some prior information, or to a value which is clearly larger than
the expected number of sources. It is also possible to try different numbers of
components and to determine a suitable value of A^ from the outcome of the
trials.
As discussed in the previous section, non-negativity restrictions can be
used for frequency-domain basis functions. With a sparse prior and non-
negativity restrictions, one has to use the projected steepest descent algo-
rithms which are discussed, e.g., by Bertsekas in [35, pp. 203-224]. Hoyer
[299], [300] proposed a non-negative sparse coding algorithm by combining
NMF and sparse coding. His algorithm used a multiplicative rule to update
B, and projected steepest descent to update G. Projected steepest descent
alone is computationally inefl[icient compared to multiplicative update rules,
for example.
In musical signal analysis, sparse coding has been used for example
by Abdallah and Plumbley [4], [5] to produce an approximate piano-roll
transcription of synthesized harpsichord music and by Virtanen [650] to tran-
scribe drums in polyphonic music signals synthesized from MIDI. Also, Blu-
mensath and Davies used a sparse prior for the gains, even though their system
was based on a different signal model [43]. The framework also enables the use
282 Tuomas Virtanen

of further assumptions. Virtanen used a cost function which included a term

that favoured the temporal continuity of gains by making large gain changes
between adjacent frames unlikely [650].

9.5 Non-Negative Matrix Factorization

As discussed in Section 9.3.2 (see p. 277), it is reasonable to restrict frequency-

domain basis functions and their gains to non-negative values. In the signal
model X ^ B G , the element-wise non-negativity of B and G alone is a
sufficient condition for the separation of sources in many cases, without an
explicit assumption of the independence of the sources.
Paatero and Tatter proposed an NMF algorithm in which the weighted
energy of the residual matrix X B G was minimized by using a least-squares
algorithm where B and G were alternatingly updated under non-negativity
restrictions [492]. More recently, Lee and Seung [399, 400] proposed NMF
algorithms which have been used in several machine learning tasks since the
algorithms are easy to implement and modify.
Lee and Seung proposed two cost functions and estimation algorithms to
obtain X ^ B G [400]. The cost functions are the square of the Euclidean
distance deuc and divergence ddiv, which are defined as

deuc(B,G) = | | X - B G | | 2 , (9.18)

and
ddiv(B, G) = Y1 D([X]fc.t, [BG]k,t), (9.19)
k,t

where the function D is defined as

D{p,q)^p\og^-p + q. (9.20)

Both cost functions are lower-bounded by zero, which is obtained only

when X = B G . It can be seen that the Euclidean distance is equal to the first
term in (9.15). Minimization of the Euclidean distance leads to a maximum
likelihood estimator for B and G in the presence of Gaussian noise. Similarly,
minimization of the divergence (9.19) leads to a maximum likelihood estima-
tor, when the observations are generated by a Poisson process with mean value
[BG]fc,, [399]. When ZkA^Wt = Efc,JBG]fe,t = 1, the divergence (9.19) is
equal to the Kullback-Leibler divergence, which is widely used as a distance
measure between probability distributions [400].
The estimation algorithms of Lee and Seung minimize the chosen cost
function by initializing the entries of B and G with random positive values,
and then by updating them iteratively using multiplicative rules. Each update
decreases the value of the cost function until the algorithm converges, i.e.,
reaches a local minimum. Usually, B and G are updated alternately.
 9 Unsupervised Learning Methods for Source Separation 283

The update rules for the EucHdean distance are given as

B ^ B.x(XG"^)./(BGG'^) (9.21)

and
G ^ G.X(B"^X)./(B'^BG), (9.22)
where .x and ./ denote the element-wise multiplication and division, respec-
tively. The update rules for the divergence are given as

and
G^G.^^&^. (9.24)

where 1 is an all-ones K-hy-T matrix, and ^ denotes the element-wise division

of matrices X and Y.
To summarize, the algorithm for NMF is as follows:

Algorithm 9.1: Non-Negative Matrix Factorization

1. Initialize each entry of B and G with the absolute values of Gaussian noise.
2. Update G using either (9.22) or (9.24) depending on the chosen cost function.
3. Update B using either (9.21) or (9.23) depending on the chosen cost function.
4. Repeat Steps (2)--(3) until the values converge.

Methods for the estimation of the number of components have not been
proposed, but all the methods suggested in Section 9.4 are applicable in NMF,
too. The multiplicative update rules have proven to be more efficient than for
example the projected steepest-descent algorithms [400], [299], [5].
NMF can be used only for a non-negative observation matrix and therefore
it is not suitable for the separation of time-domain signals. However, when
used with the magnitude or power spectrogram, the basic NMF can be used
to separate components without prior information other than the element-
wise non-negativity. In particular, factorization of the magnitude spectrogram
using the divergence often produces relatively good results. The divergence
cost of an individual observation [X.]k,t is linear as a function of the scale of
the input, since D{ap^aq) = aD(p, g) for any positive scalar a, whereas for
the Euclidean cost the dependence is quadratic. Therefore, the divergence is
more sensitive to small-energy observations.
NMF does not explicitly aim at components which are statistically in-
dependent from each other. However, it has been proved that under certain
conditions, the non-negativity restrictions are theoretically sufficient for sep-
arating statistically independent sources [525]. It has not been investigated
whether musical signals fulfill these conditions, and whether NMF implement
284 Tuomeis Virtanen

a suitable estimation algorithm. Currently, there is no comprehensive theo-

retical explanation of why NMF works so well in sound source separation.
If a mixture spectrogram is a sum of sources which have a static spectrum
with a time-varying gain, and each of them is active in at least one frame
and frequency line in which the other components are inactive, the objec-
tive function of NMF is minimized by a decomposition in which the sources
are separated perfectly. However, real-world music signals rarely fulfill these
conditions. When two or more more sources are present simultaneously at all
times, the algorithm is likely to represent them with a single component.
In the analysis of music signals, the basic NMF has been used by Smaragdis
and Brown [600], and extended versions of the algorithm have been pro-
posed for example by Virtanen [650] and Smaragdis [599]. The problem of
the large dynamic range of musical signals has been addressed e.g. by Abdal-
lah and Plumbley [5]. By assuming multiphcative gamma-distributed noise in
the power spectral domain, they derived the cost function

D(p,g) = ^ - l + log^, (9.25)

to be used instead of (9.20). Compared to the Euclidean distance (9.18) and

divergence (9.20), this distance measure is more sensitive to low-energy ob-
servations. In our simulations, however, it did not produce results as good as
the EucUdean distance or the divergence did.

9.6 Prior Information about Sources

Manual transcription of music requires a lot of prior knowledge and training.
The described separation algorithms used some general assumptions about
the sources in the core algorithms, such as independence or non-negativity,
but also other prior information on the sources is often available. For example
in the analysis of pitched musical instruments, it is known in advance that
the spectra of instruments are approximately harmonic. Unfortunately, it is
difficult to implement harmonicity restrictions in the models discussed earlier.
Prior knowledge can also be source-specific. The most common approach to
incorporate prior information about sources in the analysis is to train source-
specific basis functions in advance. Several approaches have been proposed.
The estimation is usually done in two stages, which are
1. Learn source-specific basis functions from training material, such as mono-
timbral and monophonic music. Also the characteristics of time-varying
gains can be stored, for example by modelhng their distribution.
2. Represent a polyphonic signal as a weighted sum of the basis functions of
all the instruments. Estimate the gains and keep the basis functions fixed.
It is not yet known whether automatic music transcription is possible without
any source-specific prior knowledge, but obviously this has the potential to
make the task much easier.
 9 Unsupervised Learning Methods for Source Separation 285

Several methods have been proposed for training the basis functions in
advance. The most straightforward choice is to also separate the training
signal using some of the described methods. For example, Jang and Lee [314]
used ISA to train basis functions for two sources separately. Benaroya et al.
[32] suggested the use of non-negative sparse coding, but they also tested using
the spectra of random frames of the training signal as the basis functions or
grouping similar frames to obtain the basis functions. They reported that
non-negative sparse coding and the grouping algorithm produced the best
results [32]. Gautama and Van Halle compared three different self-organizing
methods in the training of basis functions [204].
The training can be done in a more supervised manner by using a sepa-
rate set of training samples for each basis function. For example in the drum
transcription systems proposed by FitzGerald et al. [188] and Paulus and
Virtanen [505], the basis function for each drum instrument was calculated
from isolated samples of each drum. It is also possible to generate the basis
functions manually, for example so that each of them corresponds to a single
pitch. Lepain used frequency-domain harmonic combs as the basis functions,
and parameterized the rough shape of the spectrum using a slope parameter
[403]. Sha and Saul trained the basis function for each discrete fundamental
frequency using a speech database with annotated pitch [579].
In practice, it is difficult to train basis functions for all the possible sources
beforehand. An alternative is to use trained or generated basis functions which
are then adapted to the observed data. For example, Abdallah and Plumbley
initialized their non-negative sparse coding algorithm with basis functions that
consisted of harmonic spectra with a quarter-tone pitch spacing [5]. After the
initialization, the algorithm was allowed to adapt these.
Once the basis functions have been trained, the observed input signal is
represented using them. Sparse coding and non-negative matrix factorization
techniques are feasible also in this task. Usually the reconstruction error be-
tween the input signal and the model is minimized while using a small number
of active basis functions (sparseness constraint). For example, Benaroya et al.
proposed an algorithm which minimizes the energy of the reconstruction error
while restricting the gains to be non-negative and sparse [32].
If the sparseness criterion is not used, a matrix G reaching the global
minimum of the reconstruction error can be usually found rather easily. If the
gains are allowed to have negative values and the estimation criterion is the
energy of the residual, the standard least-squares solution
G = (B"^B)-^B"^X (9.26)

produces the optimal gains (assuming that the previously trained basis func-
tions are linearly independent) [339, pp. 220-226]. If the gains are restricted
to non-negative values, the least-squares solution is obtained using the non-
negative least-squares algorithm [397, p. 161]. When the basis functions,
observations, and gains are restricted to non-negative values, the global min-
imum of the divergence (9.19) between the observations and the model can
286 Tuomas Virtanen

be computed by applying the multiplicative update (9.24) iteratively [563],

[505]. Lepain minimized the sum of the absolute value of the error between
the observations and the model by using linear programming and the Simplex
algorithm [403].
The estimation of the gains can also be done in a framework which in-
creases the probability of basis functions being non-zero in consecutive frames.
For example, Vincent and Rodet used hidden Markov models (HMMs) to
model the durations of the notes [648].
It is also possible to train prior distributions for the gains. Jang and Lee
used standard ICA techniques to train time-domain basis functions for each
source separately, and modelled the probabihty distribution function of the
component gains with a generalized Gaussian distribution which is a family
of density functions of the form p(x) oc exp(|x|^) [314]. For an observed
mixture signal, the gains were estimated by maximizing their posterior prob-
ability.

9.7 Further Processing of the Components

The main motivation for separating an input signal into components is that
each component usually represents a musically meaningful entity, such as a
percussive instrument or all the equal-pitched notes of an instrument. Separa-
tion alone does not solve the transcription problem, but has the potential to
make it much easier. For example, estimation of the fundamental frequency
of an isolated sound is easier than multiple fundamental frequency estimation
in a mixture signal.

9.7.1 Associating Components with Sources

If the basis functions are estimated from a mixture signal, we do not know
which component is produced by which source. Since each source is modelled
as a sum of one or more components, we need to associate the components
to sources. There are roughly two ways to do this. In the unsupervised classi-
fication framework, component clusters are formed based on some similarity
measure, and these are interpreted as sources. Alternately, if prior informa-
tion about the sources is available, the components can be classified to sources
based on their distance to source models. Naturally, if pre-trained basis func-
tions are used for each source, the source of each basis function is known and
classification is not needed.
Pairwise dependence between the components can be used as a similarity
measure for clustering. Even in the case of ICA, which aims at maximizing the
independence of the components, some dependencies may remain because it
is possible that the input signal contains fewer independent components than
are to be separated.
 9 Unsupervised Learning Methods for Source Separation 287

Casey and Westner used the symmetric Kullback-Leibler divergence be-

tween the probability distribution functions of basis functions as a distance
measure, resulting in an independent component cross-entropy matrix (an 'ix-
egram') [73]. Dubnov proposed a distance measure derived from the higher-
order statistics of the basis functions or the gains [157]. Casey and Westner
[73] and Dubnov [157] also suggested clustering algorithms for grouping the
components into sources. These try to minimize the inter-cluster dependence
and maximize the intra-cluster dependence.
For predefined sound sources, the association can be done using pattern
recognition methods. Uhle et al. extracted acoustic features from each com-
ponent to classify them either to a drum track or to a harmonic track [634].
The features in their system included, for example, the percussiveness of the
time-varying gain, and the noise-likeness and dissonance of the spectrum. An-
other system for separating drums from polyphonic music was proposed by
Helen and Virtanen. They trained a support vector machine (SVM) using the
components extracted from a set of drum tracks and polyphonic music sig-
nals without drums. Different acoustic features were evaluated, including the
above-mentioned ones, mel-frequency cepstral coefficients, and others [282].

9.7.2 Extraction of Musical Information

The separated components are usually analysed to obtain musically important

information, such as the onset and offset times and fundamental frequency of
each component (assuming that they represent individual notes of a pitched
instrument). Naturally, the analysis can be done by synthesizing the com-
ponents and by using analysis techniques discussed elsewhere in this book.
However, the synthesis stage is usually not needed, but analysis using the ba-
sis functions and gains directly is likely to be more reliable, since the synthesis
stage may cause some artifacts.
The onset and offset times of each component n are measured from the
time-varying gains gn^ti t 1 . . .T. Ideally, a component is active when its
gain is non-zero. In practice, however, the gain may contain interference from
other sources and the activity detection has to be done with a more robust
method.
Paulus and Virtanen [505] proposed an onset detection procedure that was
derived from the psychoacoustically motivated method of Klapuri [347]. The
gains of a component were compressed, differentiated, and lowpass filtered.
In the resulting 'accent curve', all local maxima above a fixed threshold were
considered as sound onsets. For percussive sources or other instruments with
a strong attack transient, the detection can be done simply by locating local
maxima in the gain functions, as done by FitzGerald et al. [188].
The detection of sound offsets is a more difficult problem, since the am-
plitude envelope of a note can be exponentially decaying. Methods to be used
in the presented framework have not been proposed.
288 Tuomas Virtanen

There are several different possibilities for the estimation of the funda-
mental frequency of a pitched component. For example, prominent peaks can
be located from the spectrum and the two-way mismatch procedure of Maher
and Beauchamp [428] can be used, or the fundamental period can be esti-
mated from the autocorrelation function which is obtained by inverse Fourier
transforming the power spectrum. In our experiments, the enhanced auto-
correlation function proposed by Tolonen and Karjalainen [627] was found to
produce good results (see p. 253 in this volume). In practice, a component may
represent more than one pitch. This happens especially when the pitches are
always present simultaneously, as is the case in a chord, for example. No meth-
ods have been proposed to detect this situation. Whether or not a component
is pitched can be estimated, e.g., from features based on the component [634],
[282].
Some systems use fixed basis functions which correspond to certain funda-
mental frequency values [403], [579]. In this case, the fundamental frequency
of each basis function is of course known.

9.7.3 Synthesis

Synthesis of the separated components is needed at least when one wants to

listen to them, which is a convenient way to roughly evaluate the quality of the
separation. Synthesis from time-domain basis functions is straightforward: the
signal of component n in frame t is generated by multiplying the basis function
hn by the corresponding gain gn,t^ and adjacent frames are combined using
the overlap-add method where frames are multiplied by a suitable window
function, delayed, and summed.
Synthesis from frequency-domain basis functions is not as trivial. The syn-
thesis procedure includes calculation of the magnitude spectrum of a compo-
nent in each frame, estimation of the phases to obtain the complex spectrum,
and an inverse discrete Fourier transform (IDFT) to obtain the time-domain
signal. Adjacent frames are then combined using overlap-add. When magni-
tude spectra are used as the basis functions, framewise spectra are obtained
as the product of the basis function with its gain. If power spectra are used, a
square root has to be taken, and if the frequency resolution is not hnear,
additional processing has to be done to enable synthesis using the IDFT.
A few alternative methods have been proposed for the phase generation.
Using the phases of the original mixture spectrogram produces good syn-
thesis quality when the components do not overlap significantly in time and
frequency [651]. However, applying the original phases and the IDFT may pro-
duce signals which have unrealistic large values at frame boundaries, resulting
in perceptually unpleasant discontinuities when the frames are combined us-
ing overlap-add. The phase generation method proposed by Griffin and Lim
[259] has also been used in synthesis (see for example Casey [70]). The method
finds phases so that the error between the separated magnitude spectrogram
and the magnitude spectrogram of the resynthesized time-domain signal is
 9 Unsupervised Learning Methods for Source Separation 289

minimized in the least-squares sense. The method can produce good synthesis
quaUty especially for slowly varying sources with deterministic phase behav-
iour. The least-squares criterion, however, gives less importance to low-energy
partials and often leads to a degraded high-frequency content. The phase gen-
eration problem has been recently addressed by Achan et ah, who proposed
a phase generation method based on a pre-trained autoregressive model [9].

9.8 Time-Varying Components

As mentioned above, the linear model (9.1) is efficient in the analysis of music
signals since many musically meaningful entities can be rather well approxi-
mated with a fixed spectrum and a time-varying gain. However, representation
of sources with strongly time-varying spectrum requires several components,
and each fundamental frequency value produced by a pitched instrument has
to be represented with a different component. Instead of using multiple com-
ponents per source, more complex models can be constructed which allow
either a time-varying spectrum or a time-varying fundamental frequency for
each component. These are discussed in the following two subsections.

9.8.1 Time-Varying Spectra

Time-varying spectra of components can be obtained by replacing each basis

function b by a sequence of basis functions bn,r, where r = 0 . . . L 1 is the
frame index. If a frequency-domain representation is used, this means that a
static short-time spectrum of a component is replaced by a spectrogram of
length L frames.
The signal model for one component can be formulated as a convolution
between its spectrogram and time-varying gain. The model for a mixture
spectrum of A^ components is given by
N L-1
Xt ^ ^ ^ hn,T9n,t-T' (9.27)
n=lT=0

The model can be interpreted so that each component n consists of repetitions

of an event which has a spectrogram bn,r,7- = 0...L 1. Each non-zero value
of the time-varying gain gn,t denotes an onset of the event and the value of
the gain gives the scaling factor of each repetition. A simple two-note example
is illustrated in Fig. 9.4.
The parameters of the convolutive model (9.27) can be estimated using
methods extended from NMF and sparse coding. In these, the reconstruction
error between the model and the observations is minimized, while restricting
the parameters to be entry-wise non-negative. Also favouring sparse gains is
clearly reasonable, since real-world sound events set on in a small number of
290 Tuomas Virtanen

5000
- -
^4000
o '
3000 ' ~
o
2000
t ..,
^ 1000
" 1
E0.5T " 1.5 2 2.5
Time (seconds)

F i g . 9.4. An example of the convolutive model (9.27) which allows time-varying

components. The mixture spectrogram (upper left panel) contains the notes C # 6
and F # 6 of the acoustic guitar, first played separately and then together. The upper
right panels illustrate the learned note spectrograms and the lower panel shows their
time-varying gains. In the gains, an impulse corresponds to the onset of a note. The
components were estimated using a modified version of the algorithm proposed by
Smaragdis in [599]. In the case of more complex signals, it is difficult to obtain such
clear impulses.

frames only. Virtanen [651] proposed an algorithm which is based on non-

negative sparse coding, whereas t h a t of Smaragdis [599] aims at minimizing
the divergence between the observation and the model while constraining non-
negativity.
Arbitrarily long durations L may not be used if the basis functions are
estimated from a mixture signal. W h e n NL > T, the input spectrogram can
be represented perfectly as a sum of concatenated event spectrograms (without
separation). Meaningful sources are likely to be separated only when NL <C T.
In other words, estimation of several components with large L requires long
input signals.
In addition, the method proposed by Blumensath and Davies [43] can be
formulated using (9.27). Their objective was to find sparse and shift-invariant
decompositions of a signal in the time domain. Their model allows an event
to begin at any time with one sample accuracy which makes the number
of free parameters in the model large. To reduce the dimensionality of the
problem, Blumensath and Davies proposed an algorithm which carried out
 9 Unsupervised Learning Methods for Source Separation 291

the optimization in a subspace of the parameters. They also included a sparse

prior for the gains.

9.8.2 Time-Varying Fundamental Frequencies

In some cases, it is desirable to use a model which can represent different

pitch values of an instrument with a single component. For example, in the
case where a note with a certain pitch is present only during a short time,
separating it from co-occurring sources is difficult. However, if other notes of
the source with adjacent pitch values can be utilized, the estimation becomes
more reliable.
Varying fundamental frequencies are difficult to model using time-domain
basis functions or frequency-domain basis functions with linear frequency reso-
lution. This is because changing the fundamental frequency of a basis function
is a non-linear operation which is difficult to implement in practice: if the fun-
damental frequency is multiplied by a factor 7, the frequencies of the harmonic
components are also multiplied by 7; this can be viewed as a stretching of the
spectrum. For an arbitrary value of 7, the stretching is difficult to perform on
a discrete linear frequency resolution, at least using a simple operator which
could be used in the unsupervised learning framework. The same holds as well
for time-domain basis functions.
A logarithmic spacing of frequency bins makes it easier to represent varying
fundamental frequencies. A logarithmic scale consists of discrete frequencies
fref/?'^"^, where k = 1... K is the discrete frequency index, /? > 1 is the ratio
between adjacent frequency bins, and fref is a reference frequency in Hertz
which can be selected arbitrarily. For example, (3 = v ^ produces a frequency
scale where the spacing between the frequencies is one semitone.
On the logarithmic scale, the spacing of the partials of a harmonic sound
is independent of its fundamental frequency. For fundamental frequency /o,
the overtone frequencies of a perfectly harmonic sound are m/o, where m > 0
is an integer. On the logarithmic scale, the corresponding frequency indices
are k = log^(m) -h log^(/o/fref), and thus the fundamental frequency affects
only the offset log^(/o/fref), not the intervals between the harmonics.
Given the spectrum X{k) of a harmonic sound with fundamental frequency
/o, a fundamental frequency multiplication 7/0 can be implemented simply
as a translation X{k) = X{k 6), where S is given by S = log^ 7. Compared
with the stretching of the spectrum, this is usually easier to implement.
The estimation of harmonic spectra and their translations can be done
adaptively by fitting a model onto the observations.'* However, this is diffi-
cult for an unknown number of sounds and fundamental frequencies, since the
reconstruction error as a function of translation 6 has several local minima

^This approach is related to the fundamental frequency estimation method of

Brown, who calculated the cross-correlation between an input spectrum and a single
harmonic template on the logarithmic frequency scale [54].
292 Tuomas Virtanen

at harmonic intervals, which makes the optimization procedure Hkely to be-

come stuck in a local minimum far from the global optimum. A more feasible
parameterization allows each component to have several active fundamental
frequencies in each frame, the amount of which is to be estimated. This means
that each time-varying gain gn^t is replaced by gains gn,t,zi where 2; = 0 , . . . , Z
is a frequency-shift index and Z is the maximum allowed shift. The gain gn,t,z
describes the amount of the n^^ component in frame f at a fundamental fre-
quency which is obtained by translating the fundamental frequency of basis
function b by z indices.
The size of the shift z depends on the frequency resolution. For example, if
48 frequency lines within each octave are used (/? = ^ 2 ) , z 4: corresponds to
a shift of one semitone. For simplicity, the model is formulated to allow shifts
only to higher frequencies, but it can be formulated to allow both negative
and positive shifts, too.
A vector gri,t [^n,t,0 5' ' 19n,t,z] is used to denote the gains of compo-
nent n in frame t. The model can be formulated as
N
xt^^bn*g^,,, t = l...T, (9.28)
n=l
where * denotes a convolution operator, defined between vectors as
z
y = bn * gn,t ^yk = ^ bn,k-z9n,t,z^ k = l...K. (9.29)
z=0
Figure 9.5 shows the basis function and gains estimated from the example
signal in Fig. 9.1. In general, the parameters can be estimated by fitting
the model to observations with certain restrictions, such as non-negativity or
sparseness. Algorithms for this purpose can be derived by extending those
used in NMF and sparse coding. Here we present an extension of NMF, where
the parameters are estimated by minimizing the divergence (9.19) between
the observations X and the model (9.28), while restricting the gains and basis
functions to be non-negative.
The elements of gn,t and b^ are initialized with random values and then
updated iteratively until the values converge. To simplify the notation, let us
denote the model with current parameter estimates by v^ = J2n=i ^^ * ^^^^^^^
t = 1 . . . T. The update rule for the gains is given as

gn,t-gn,..X-^, (9.30)
bn 1
where 1 is a ivT-length vector of ones and denotes the correlation of vec-
tors, defined for real-valued vectors bn and y as g = b^ ^ y <^ Pz =
Sfc=i ^n,fc2/fc+25 z = 0 , . . . , Z. The update rule for the basis functions is
given as
 9 Unsupervised Learning Methods for Source Separation 293

bn ^ b.x (9.31)

T h e overall optimization algorithm for non-negative matrix deconvolution

is as follows:

Algorithm 9.2: N o n - N e g a t i v e Matrix Deconvolution

1. Initialize each gn,t and bn with the absolute values of Gaussian noise.
2. Calculate v* = ^ ^ ^ ^ bn * gn,t for each t=l...T.
3. Update each gn,t using (9.30).
4. Calculate Vt as in Step 2.
5. Update each bn using (9.31). Repeat Steps (2)-(5) until the values converge.

T h e algorithm produces good results if the number of sources is small, b u t

for multiple sources and more complex signals, it is difficult to get as good

50
'ui V I 10
(D
C k
o
40 -
w
Hi
0)
c
(Q
0-30 . m-10
M i mm.
*;
x:
"D
(0
O -20
f, 20 p. Q.
E
3
V^S:K ap:> <
H-

ca -30
Z 10 -
CD v>
\r
CO
o
c
-40
13 ^
u- 0

-50
1 2 3 10 10
Time (seconds) Frequency (Hertz)

Fig. 9.5. Illustration of the time-varying gains (left) and the basis function (right)
of a component that was estimated from the example signal in Fig. 9.1 containing
a diatonic scale and a C major chord. On the left, the intensity of the image rep-
resents the value of the gain at each fundamental frequency shift and frame index.
Here the fundamental frequencies of the notes can be seen more clearly than from
the spectrogram of Fig. 9.1. The parameters were estimated using the algorithm
proposed in this section.
294 Tuomas Virtanen

results as those illustrated in Fig. 9.5. The model allows all the fundamental
frequencies within the range z = 0.. .Z tohe active simultaneously, thus, it is
not restrictive enough. For example, the algorithm may model a non-harmonic
drum spectrum by using a harmonic basis function shifted to multiple adjacent
fundamental frequencies. Ideally, this could be solved by restricting the gains
to be sparse, but the sparseness criterion complicates the optimization.
In principle, it is possible to combine time-varying spectra and time-
varying fundamental frequencies into the same model, but this further in-
creases the number of free parameters so that it can be difficult to obtain
good separation results.
When shifting the harmonic structure of the spectrum, the formant struc-
ture becomes shifted, too. Therefore, representing time-varying pitch by trans-
lating the basis function is appropriate only for nearby pitch values. It is
unlikely that the whole fundamental frequency range of an instrument could
be modelled by shifting a single basis function.

9.9 Evaluation of the Separation Quality

A necessary condition for the development of source separation methods is
the ability to measure the quality of their results. In general, the separation
quality can be measured by calculating the error between the separated sig-
nals and reference sources, or by listening to the separated signals. In the case
that separation is used as a pre-processing step for automatic music tran-
scription, the quality should be judged according to the final application, i.e.,
the transcription accuracy.
Performance measures for audio source separation tasks have been dis-
cussed, e.g., by Gribonval et al. [258]. They proposed measures estimating the
amount of interference from other sources and the distortion caused by the
separation algorithm. Many authors have used the signal-to-distortion ratio
(SDR) as a simple measure to summarize the quality. This is defined in deci-
bels as
SDR[dB] = 1 0 1 o g , ^ T ^ j s(t)'^
|^^l_, (9.32)

where s{t) is a reference signal of the source before mixing, and s{t) is the sep-
arated signal. In the separation of music signals, Jang and Lee [314] reported
average SDR of 9.6 dB for an ISA-based algorithm which trains basis func-
tions separately for each source. Helen and Virtanen [282] reported average
SDR of 6.4 dB for NMF in the separation of drums and polyphonic harmonic
track, and a clearly lower performance (SDR below 0 dB) for ISA.
In practice, quantitative evaluation of the separation quality requires that
reference signals, i.e., the original signals s{t) before mixing, be available. In
the case of real-world music signals, it is difficult to obtain the tracks of each
individual source instrument and, therefore, synthesized material is often used.
 9 Unsupervised Learning Methods for Source Separation 295

Generating test signals for this purpose is not a trivial task. For example, ma-
terial generated using a software synthesizer may produce misleading results
for algorithms which learn structures from the data, since many synthesiz-
ers produces notes which are identical at each repetition. In the case that
source separation is a part of a music transcription system, quality evaluation
requires that audio signals with an accurate reference notation are available
(see Chapter 11, p. 355). Large-scale comparisons of different separation al-
gorithms for music transcription have not been made.

9.10 S u m m a r y and Discussion

The algorithms presented in this chapter show that rather simple principles
can be used to learn and separate sources from music signals in an unsuper-
vised manner. Individual musical sounds can usually be modelled quite well
using a fixed spectrum with time-varying gain, which enables the use of ICA,
sparse coding, and NMF algorithms for their separation. Actually, all the al-
gorithms based on the linear model (9.4) can be viewed as performing matrix
factorization; the factorization criteria are just different.
The simplicity of the additive model makes it relatively easy to extend
and modify it, along with the presented algorithms. However, a challenge
with the presented methods is that it is difficult to incorporate some types of
restrictions for the sources. For example, it is difficult to restrict the sources
to be harmonic if they are learned from the mixture signal.
Compared to other approaches towards monaural sound source separa-
tion, the unsupervised methods discussed in this chapter enable a relatively
good separation qualityalthough it should be noted that the performance
in general is still very limited. A strength of the presented methods is their
scalability: the methods can be used for arbitrarily complex material. In the
case of simple monophonic signals, they can be used to separate individual
notes, and in complex polyphonic material, the algorithms can extract larger
repeating entities, such as chords. Some of the algorithms, for example NMF
using the magnitude spectrogram representation, are quite easy to imple-
ment. The computational complexity of the presented methods may restrict
their applicability if the number of components is large or the target signal is
long.
Large-scale evaluations of the described algorithms on real-world poly-
phonic music recordings have not been presented. Most published results use
a small set of test material and the results are not comparable with each
other. Although conclusive evaluation data are not available, a preliminary
experience from our simulations has been that NMF (or sparse coding with
non-negativity restrictions) often produces better results than ISA. It was
also noticed that prior information about sources can improve the separation
quality significantly. Incorporating higher-level models into the optimization
296 Tuomas Virtanen

algorithms is a big challenge, but will presumably lead to better results. Con-
trary to the general view held by most researchers less than 10 years ago,
unsupervised learning has proven to be applicable for the analysis of real-
world music signals, and the area is still developing rapidly.
 Part IV

Entire Systems, Acoustic and Musicological

Modelling
10
Auditory Scene Analysis in Music Signals

Kunio Kashino

NTT Communication Science Laboratories,

Nippon Telegraph and Telephone Corporation
3-1, Morinosato-Wakamiya, Atsugi, 243-0198, Japan
kunioQeye.brl.ntt.co.jp

10.1 Introduction

This chapter discusses work done in the area of music scene analysis (MSA).
Generally, scene analysis is viewed as the transformation of information from
a sensory input (physical entity) into concepts (psychological or perceptual
entities). Therefore, MSA is defined as a process that converts an audio signal
into musical concepts such as notes, chords, beats, and rhythms. Related tasks
include music transcription, pitch tracking, and beat tracking; however, this
chapter focuses on the auditory scene analysis (ASA) related aspect of this
process and does not explore the issues of pitch and beat tracking. An impor-
tant idea related to this is the distinction between physical and perceptual
sounds, as explained in Section 10.1.3 below.

10.1.1 Scene Analysis

We are exposed to various physical stimuli in our daily life. Our ears and eyes
receive acoustic and optical stimuli, respectively. These stimuli originate in
specific events or states. For example, when a ball hits a wall, vibrations in
both the ball and the wall resulting from the impact travel through the air,
and the air vibration arrives at our ears. We then understand that something
like a ball has hit a hard surface such as a wall.
Understanding physical stimuli is an everyday experience. However, it
poses an important question. An event such as a ball hitting a wall and causing
physical phenomena such as air vibration is a natural process. However, how
can we determine the events from the received physical phenomena? This is
an inversion of the natural process, and therefore the solution to this problem
is non-trivial.
Generally, a task that consists of recognizing an event or status from phys-
ical stimuli is called scene analysis, and scene analysis problems were first
investigated in the visual domain.
300 Kunio Kashino

In the 1920s, psychology researchers noticed the problem involved in de-

termining how humans organize perceptual entities from visual stimuli. Their
series of studies was referred to as 'Gestalt psychology'. In the 1960s and
1970s, pioneering work was undertaken on computer vision. For example, the
problem of describing a simple scene comprising a set of building blocks with
a simpler set of building blocks such as triangular and rectangular prisms was
considered. Various operators for feature extraction were also proposed. In the
1970s and 1980s, many researchers addressed the problem of building generic
scene analysis systems. In particular, a processing architecture comprising
multiple processing modules and a common space for exchanging information
among the modules was applied to scene analysis tasks. Such an architecture
is called the blackboard architecture.
In the 1980s, Marr proposed a computational approach. He defined the vi-
sion problem in relation to the information processing needed to describe the
visual world, and distinguished three levels of the task: computational theory,
representation and algorithm, and hardware implementation [438]. Along with
this computational approach, many researchers reported physical or mathe-
matical formulations. Examples include work on 'shape-from-X' problems that
address shape recovery by using various clues, denoted by X, such as motion,
shading, and multiple images.
There is still a large amount of research being reported that relates to
visual scene analysis. The amount of work on visual scene analysis from a
generic perspective has decreased in recent years, and most of this work is
dedicated to specific targets such as information retrieval, robot navigation,
motion analysis, and encoding.
The history of auditory scene analysis is rather short compared with that
of visual scene analysis, although the 'cocktail party problem'^ was raised as
early as the 1950s [86], [466]. In auditory psychophysics, various phenomena
have been found that show that our auditory system has complex, adaptive,
and active functions [662], [475], [270], [49], and auditory versions of Gestalt
principles have been accumulated with respect to the way in which humans
organize complex auditory stimuli into 'auditory streams' that correspond to
distinct perceptual entities. However, from an engineering point of view, there
are still many important problems to overcome in the auditory scene analysis
field.

10.1.2 Music as an Auditory Scene

Music is a good domain for considering the auditory scene analysis prob-
lem not only from a cognitive perspective [36] but also from an engineering
viewpoint. We use the term 'music scene analysis' to refer to auditory scene
analysis for music.

^The cocktail party problem refers to the task of following the discussion of one's
neighbours in a situation where lots of other sound sources are present, too.
 10 Auditory Scene Analysis in Music Signals 301

The nature of music provides the first reason for considering it as an audi-
tory scene. The overlapping of tones is a fundamental element of music. With
the exception of a solo performance by a single instrument or voice, music
usually comprises multiple simultaneous sounds played by single or multiple
musical instruments. As humans we can appreciate such sound mixtures, im-
plying that we recognize what is happening to a certain extent. Only a trained
person is capable of transcribing music, but ordinary people can recognize a
vocal line or the principal accompaniment when they listen to pop music.
When we hear music played by a flute and a piano, it is easy to distinguish
the two instruments.
The second reason is the usefulness of the prospective applications. Cur-
rently, computers are not as good as humans at recognizing multiple simul-
taneous sounds. However, if computers are developed with this capability,
various useful systems will be realized including automatic music transcrip-
tion systems and automatic music indexing systems for unlabelled music
archives.
A research topic closely related to music scene analysis is automatic music
transcription [610] and pitch tracking. As already introduced in Chapter 1,
an automatic music transcription system was reported as long ago as the mid
1970s [521]. From then until the mid 1980s, several systems were built that
mainly targeted the transcription of monophonic melodies such as singing, or
simple polyphonic music such as guitar duets [483], [542]. The main method-
ology employed in such work involved signal processing techniques such as the
fast Fourier transform (FFT). This period can be considered as the pioneering
era of music transcription. Although the systems targeted rather simple com-
positions, various problems were identified such as frequency and temporal
fluctuations.
From the mid 1980s to the mid 1990s, the main target moved from mono-
phonic music to rather complicated polyphonic music, such as piano composi-
tions. With such signals, even determining the number of simultaneous notes
is a hard task let alone extracting fundamental frequencies for each note. To
overcome this problem, researchers pointed out that knowledge is required
for such transcription [470], [81], [80]. The main methodology consisted of in-
tegrating symbolic knowledge and signal processing. For example, Katayose
et al. built a rule-based automatic music transcription system for multiple
simultaneous-note performances by a single instrument [335]. The system com-
prises a control module, a processing module, and a music analysis module.
The control module is an inference engine performing rule-based reasoning
and invoking the processing module that extracts fundamental frequencies
and beat times. The music analysis module analyses musical characteristics
such as melody, rhythm, chord transitions, and keys, and then its results are
fed back to the control and processing modules. This type of approach to some
degree parallels the methodology of visual scene analysis in use at that time
as mentioned above. This period can be viewed as the system-oriented era of
music transcription.
302 Kunio Kashino

At that time, in the artificial inteUigence field, it was recognized that there
was a bottleneck as regards knowledge acquisition for rule-based inference
engines. This means that it was diflicult to prepare all the knowledge required
for these systems, and the systems tended to fail to work properly when they
faced a situation not dealt with by the installed knowledge.
Since the mid 1990s, a lot of research has targeted polyphonic music played
by multiple musical instruments in, for example, orchestras or commercial pop
music [611], [350], [353]. Music played on multiple instruments is even harder
to transcribe than music played on a single instrument, mainly because there
are fewer constraints on the instruments or because we have less prior knowl-
edge. That is, we must estimate the most plausible transcription under highly
ambiguous and uncertain situations. Researchers noticed that it is hard to
deal with such problems solely by the rule-based approach, and this require-
ment has naturally inspired researchers to apply probabilistic modelling. We
call this period the model-oriented era of music transcription.
As is widely recognized, music usually has simultaneous and temporally hi-
erarchical structures. For example, multiple simultaneous notes form a chord,
and a sequence of notes across multiple bars form a phrase. Such structures
are naturally incorporated in probabilistic models.
This is the point that distinguishes music transcription from music scene
analysis. The object of music transcription is to create a score from a musical
audio signal. On the other hand, the object of music scene analysis is to
recover hierarchical structures and describe the auditory entities encoded in
the structures from a musical audio signal. The recognition of structures is
not a prerequisite for creating a score. However, a score can be produced once
a complete music structure has been obtained. Prom this viewpoint, music
transcription can be a specific instance of music scene analysis.
The present author used the term music scene analysis in this sense, and
proposed a music scene analysis system based on a probabilistic model in the
mid 1990s [332]. We defined the problem as the estimation of the posterior
probability distribution given an input audio signal and a set of prior knowl-
edge encoded in internal models. Recently, other probabilistic approaches to-
ward music analysis tasks have been emerging. For example. Goto highlighted
a sub-symbolic^ aspect in music scene analysis and specifically termed it music
scene description [228]. As discussed in Chapter 11, his descriptions of, for ex-
ample, predominant fundamental frequencies and beat times correspond to a
primal level of a music scene representation structure, which is very important
to address.

10.1.3 Perceptual Sounds and Their Structure

To specify the problem of music scene analysis a little more precisely, here
we introduce the terms 'physical sound' and 'perceptual sound'. A physical
^A sub-symbolic description here means description that involves continuous
quantity.
 10 Auditory Scene Analysis in Music Signals 303
Chord C
chord

piano E

component

^ ^ : perceptual sound

Fig. 10.1. An example snapshot of perceptual sounds in music.

sound means the vibration of the medium itself. It is a physical entity. On

the other hand, there is another aspect to sound, namely perceptual sound,
which is a psychological entity that corresponds to what human ear perceives
as 'one' sound.
Suppose we are listening to ensemble music through a monaural speaker. In
this case, there is one physical sound source. However, we will hear multiple
perceptual sounds produced by multiple musical instruments. That is, the
number of perceptual sound sources is greater than one.
In this sense, the concept of a perceptual sound is similar to that of an
'auditory stream', as used by Bregman [49]. However, a perceptual sound is
not necessarily a stream but an entity in an intrinsically hierarchical structure.
Acoustic energy must be organized in a structure if we are to understand the
scene. In reality, what humans hear as one sound depends on time, place,
occasion, and even attention. For example, when we listen to music, we hear
multiple levels of perceptual sounds; sometimes we hear notes or a melody,
and at other times we hear chords. As another example, when we are waiting
for someone on a busy street, we may sometimes hear all the street noises
as one sound, but sometimes we may hear car noises or people's footsteps as
one sound. If we are specifically interested, we can focus on an individual car
noise, and furthermore we may hear the specific car's engine noise, wind noise,
or road noise individually.
Figure 10.1 shows an example snapshot of perceptual sounds in a music
performance. In Fig. 10.1, a (perceptual) component corresponds to a (phys-
ical) frequency component, which is a continuous time-frequency region on a
sound spectrogram. Figure 10.1 only shows a snapshot, but in fact there is
also a time structure.
Table 10.1 lists the meanings of the terms used in this chapter.
304 Kunio Kashino

Table 10.1. Terminology

Term Meaning
perceptual sound description corresponding to a sound that humans
perceive as one
auditory scene analysis construction, restoration, and organization of percep-
tual sounds to build a description of the world
music scene analysis auditory scene analysis for musical audio signals

10.1.4 Problem Formulation

For a music scene analysis task, we should use multiple constraints or pieces
of information stored in advance. For example, in Western tonal music, a
sequence of chords does not appear at random but exhibits certain statistical
characteristics, and multiple frequency components whose frequencies have a
harmonic relationship tend to arise from 'one sound'.
Thus, the problem can be formulated as an a posteriori estimation. Let
H he 8i set of random variables corresponding to internal states or perceptual
sounds to be modelled, and the observation be x. Then, the task is generally
written as
H = argmaxP(i7|x) = arg max F(x | if )P(i^). (10.1)
H H
This is a Bayesian estimation of posteriors. Since it is very hard to calculate
this in a general form, we must impose a structure on H. That is, when
some elements of H can be considered to be independent, the calculations can
take advantage of that fact. This point will be discussed later in this chapter.

10.2 Strategy for Music Scene Analysis

This section discusses the clues that may be used for music scene analysis.
First, we review work on sound source separation. We then look at cues or
information sources that have been utilized to associate time-frequency com-
ponents to a same or to a different perceptual sound.

10.2.1 Sound Source Separation

From an engineering point of view, topics related to auditory scene analysis or

music scene analysis include sound source separation. Since the aim of sound
source separation is to separate the source signals, it is clearly a different task
from auditory or music scene analysis where the goal is to obtain perceptual
description of the content. However, one of the major difficulties in scene
analysis is that we have to deal with mixtures of sounds. Therefore, if we can
separate signals corresponding to the physical sound sources, it may assist
scene analysis. To clarify the position of the music scene analysis problem,
 10 Auditory Scene Analysis in Music Signals 305

here we look at sound source separation, specifically from the viewpoint of

applicability to music signals.
There are two kinds of problems considered in sound source separation:
one assumes multiple input channels, and the other assumes one input channel
(monaural). In the former case, it is usually further assumed that no prior
knowledge is available about the source characteristics or mixing conditions.
As a result, this is called blind source separation [63].
A powerful methodology for dealing with the blind source separation prob-
lem is independent component analysis (ICA), as described in Chapter 9. The
basic idea is as follows. First, it is assumed that we observe linear combinations
of source waveforms:
y(n) = Ax(n), (10.2)
where the vector y(n) consists of the observed signals at time n, A is a time-
invariant mixing matrix, and the vector x(n) represents the source signals at
time n. Then, the problem is to estimate A~^. ICA is a method for estimat-
ing A~^ so that the source signals x(n) become as independent as possible
from each other. Standard ICA methods require that the number of observed
signals be larger or equal to the number of sources. In recent years, many re-
searchers have been interested in sound source separation by ICA under more
realistic conditions such as convolutive mixtures or cases where there are fewer
microphones (i.e., observed signals y(n)) than sound sources [28], [401].
For sound source separation with music signals, it is desirable for the
separation to work even for one-channel or two-channel (monaural or stereo)
signals [554], [73]. However, these cases require further research, as explored
in Chapter 9.
Research on sound source separation for monaural inputs has been con-
ducted since the 1970s. The most widely used approach is the selective
resynthesis of frequency components. Specifically, selection by employing the
harmonic frequency relationships as proposed by Parsons [498] has been widely
used. This approach is based on prior knowledge of the sound sources, namely
that sounds often consist of a fundamental frequency component and its over-
tones whose frequencies are integer multiples of the fundamental frequency. In
reality, sounds such as voiced speech or pitched musical sounds consist of ap-
proximately harmonic components and, in such cases, the harmonic selection
and resynthesis achieves reasonably good separation [482].
The selective resynthesis approach was mostly based on harmonic rela-
tions until around 1990. However, in the last two decades, rapid progress has
been made on psychophysical research designed to clarify the mechanisms
whereby humans recognize sounds. This has encouraged research efforts even
from an engineering viewpoint. Specifically, a lot of work has tried to incorpo-
rate various clues in addition to harmonic relations for component selection,
i.e., for associating components to a same source. The clues considered in
such work have included: (1) common onset and offset, (2) common ampli-
tude and frequency modulation, (3) estimated spatial locations, and (4) the
306 Kunio Kashino

Table 10.2. Clues for the integration of simultaneous frequency components

Feature of frequency components Promotes
harmonicity fusion
harmonic mistuning segregation
onset synchrony fusion
onset asynchrony segregation
offset synchrony fusion
offset asynchrony segregation
common amplitude modulation fusion
common frequency modulation fusion
spatial location fusion/segregation
timbre similarity/dissimilarity fusion/segregation

temporal and spectral proximity of frequency components. A major impetus

behind the idea of using such auditory characteristics as clues for component
selection or grouping was the accumulation of psychological findings, such as
the ones introduced by Bregman [49]. Based on the title of his book, Auditory
Scene Analysis, a series of engineering studies in this field was referred to as
'computational auditory scene analysis' [485].^
Here, we briefly introduce the clues that humans use (see also [475], [49],
[108], [98] for details). It is known that certain types of frequency component
behaviour can promote the perceptual fusion or segregation of those compo-
nents, as listed in Table 10.2.
Many authors took these clues into account in their attempts to achieve
toward sound source separation. Cooke [97] and Brown [52] incorporated psy-
chophysical and physiological knowledge into their models. An input signal is
analysed by a gammatone filterbank that simulates the frequency selectivity
of the inner ear (see Chapter 8 for more details on auditory models). Then,
the outputs of the filters are processed by Meddis's hair cell model to esti-
mate the neural activity in the auditory nerve. Autocorrelation on the model
output serves as a time-frequency energy representation. The system then ex-
tracts frequency components based on the cross-channel correlations in the
energy representation. Onset and offset synchrony and the harmonic relations
of the frequency components are used for grouping these components. Each
of the resulting groups is considered to be a single sound, such as a speech,
and resynthesized as a separate signal. They also applied their model to music
signals [53].
Mellinger [460] also introduced a physiologically and psychologically mo-
tivated model of sound source separation. He extracted and utilized common

^This terminology is slightly misleading. Originally, scene analysis meant creat-

ing a descriptive representation from signals. However, a lot of work on computa-
tional auditory scene analysis appears to address signal separation rather than scene
description.
 10 Auditory Scene Analysis in Music Signals 307

frequency modulations of frequency components and tested the separation

performance with music signals.
Nakatani [481], [480] reported a sound stream segregation system based on
harmonic relations and spatial locations. Their system employs a multi-agent
architecture and comprises three processing modules, namely a generator, a
tracer, and a monitor. The generator generates a tracer when harmonically
related and spatially localized frequency components are detected. The tracer
tracks the harmonic components as long as they continue. The monitor con-
trols the activities of the tracers to maintain global consistency. Then, the
source-separated signals are presented to an automatic speech recognition
system. They reported a significant improvement in the recognition accuracy
compared with using the mixture signal directly.
In addition, work has also been reported based on mathematical formula-
tions of the above-mentioned grouping rules [6], [636].
When considering sound source separation for music signals, it is important
to note that frequency components in music tend to overlap other components.
This is not limited to cases where the notes are exactly one octave apart in
frequency. For example. Fig. 10.2 is the beginning part of a piece entitled
'Drei Duos' composed by Beethoven. The integral multiple of the fundamental
frequency of the notes marked by x is the same as the fundamental frequency
of a note existing at the same time in another part. Since music is often
composed with notes at harmonic intervals, almost complete overlapping of
frequency components frequently occurs even in a two-part composition as in
this example. This means that the performance of the frequency component
selection approach will be limited.
As described above, sound source separation for music signals will be more
difScult than for other audio signals because the desired number of input chan-
nels is one or two, and the frequency components often overlap completely
[309], [653]. This makes it hard to decompose the music transcription or the
scene analysis problem into sound source separation and recognition of the sep-
arated signal. Therefore, many researchers have addressed the recognition or
analysis of a sound mixture as it is, without a prior source separation process.

10.2.2 B o t t o m - U p Clues for Music Scene Analysis

As mentioned above, researchers have tried to incorporate the clues for human
auditory scene analysis to sound source separation. It is also important to
consider such psychological findings in music scene analysis tasks, because a
perceptual sound is a psychological entity rather than a physical entity.
The bottom-up process of music scene analysis can be considered to be a
clustering of frequency components. The idea of clustering based on such cues
as harmonicity and synchrony has been employed by many authors for the
bottom-up processing of auditory scene analysis methods [53], [460], [171].
To formulate the clustering, it is important to explore quantitative dis-
tance measures. There are at least two approaches to this problem: one is
308 Kunio Kashino

I mrWm^

(3) (4) (3)

rt
(6) (4) (^ (5) XXX
X X X X

w^ ^ JJf ^
Fig. 10.2. Overlapping components in a composition: x indicates the notes whose
fundamental frequencies are an integral fraction of those of the notes in the other
part. The number shows the ratio of these fundamental frequencies.

Distance(V)

To be Used in the
Clustering for
Sound Formation
Frequency
Components Integration

Evaluation of Features
Fig. 10.3. The evaluation-integration model of sound formation.

to determine the measure directly from a psychometric function obtained for

simplified stimuli, and the other is to learn the measure from many realistic
sound samples. As an example, here we review an attempt using the former
approach. Kashino et al. proposed the evaluation-integration model for the
clustering for note formation [334]. Among the clues listed in Table 10.2, it
focused on the harmonic mistuning and onset asynchrony.
As shown in Fig. 10.3, the model comprises two sequential steps. The first
step involves independent evaluations of multiple features, and the second
step integrates the results of these. Here, the evaluation means determining a
distance measure between the clusters and the frequency components.
They determined the model parameters in three kinds of psychoacoustic
experiments using stimuli comprising two frequency components. The ex-
periments involved measuring: (1) the probability of segregation caused by
harmonic mistuning, (2) the probability of segregation caused by onset
asynchrony, and (3) the probability of segregation caused by both harmonic
mistuning and onset asynchrony.
 10 Auditory Scene Analysis in Music Signals 309

Two simultaneous frequency components are heard as one sound when they
are harmonically related, but they tend to be heard as separate sounds when
they are mistuned [476], [277]. In the first experiment, three subjects were
presented with stimuli with various degree of mistuning in a random order
and asked to choose whether each stimulus was one sound or two sounds.
Then, a linear model that measures segregation probability was determined
by the least squares fitting of the experimental results. The probability of
segregation Ch was given by

u, p_ < li < 0 ,
P-
Ch{u) = < ^ n^ / (10-3)

1 otherwise.

Here p+ and p- are parameters and u denotes the degree of mistuning in

percent. That is,
w= ( ^ - l ] X 100, (10.4)

where / i and /2 (/i < /2) are the frequencies of the components.
An experimental summary is shown in Fig. 10.4. The horizontal axis n in
Fig. 10.4 is given by
^,|log/2-log(2/.)|
log 1.005 ^ ^
No significant difference was found for p^ and p- and they are therefore not
distinguished in Fig. 10.4.
Two harmonic frequency components are heard as one sound when they
start simultaneously, but they tend to be heard as separate sounds when they
are asynchronous. In the second experiment, the segregation probability was
measured as a function of onset time/gradient difference, and a linear model
was obtained by least squares fitting. The probability of segregation CQ was
given by the following equation.
1

"^ (10.6)
otherwise.
5>0.

Here, S is the area of the region surrounded by the amplitude envelopes of two
frequency components projected onto the time-amplitude plane, as shown in
Fig. 10.5. In this experiment, the amplitudes of frequency components after
the onset part were chosen to be the same. The parameter Sp is given by

a b
Sp^-^ + - +c, (10.7)
/i 9i
310 Kunio Kashino
Separation p = 2.60 o.53[%] p = 2.60 0.53[%]
[%]

100

80

60

40

20

3 2 1
Downward Mistuning
... Approxinnation

Fig. 10.4. Evaluation of harmonic mistuning.

Amplitude
[dB] t

time
Fig. 10.5. An onset asynchrony model.

where / i and gi are the frequency and the onset gradient of the earlier fre-
quency component, respectively. The parameters a, 6, and c were obtained by
regression analysis, leading to the values a = 250, b = 1.11, and c = 0.317. An
example of the relation between the model and experimental results is shown
in Fig. 10.6.
The two kinds of feature evaluations are then integrated by

m{u,S) = l-{l-Ch{u)){l-Co{S)). (10.8)

where m{u,S) represents the probability of segregation. Equation (10.8) was

obtained using Dempster's rule of combination [580].
 10 Auditory Scene Analysis in Music Signals 311
Separation [%]

100

\ =

J
80
,
60
V
\i
40
\\ /

20 \
1/
0
-100 -50 0 50 100 [ms]
Relative Onset Time
of 400Hz Component
200Hz (5dB/ms) + 400Hz(ldB/ms)
Approximation

Fig. 10.6. An evaluation example of onset asynchrony.

In the third experiment, correlations were calculated between m given by

(10.8) and the experimental results when there was both harmonic mistuning
and onset asynchrony. The correlation values when the fundamental frequen-
cies were 200 Hz (i.e., sounds comprising around 200 Hz and around 400 Hz
components) and 1000 Hz (i.e., sounds comprising around 1000 Hz and around
2000 Hz components) were 0.84 and 0.75, respectively.
Equation (10.8) is the model when there are only two frequency com-
ponents. Kashino et al. introduced an approximation, and used m in (10.8)
as a distance measure for the clustering of frequency components even for
more realistic conditions [334]. The clustering algorithm, performed for each
processing window, was as follows.
1. Find the frequency component Fi that has the lowest frequency and let
it be the cluster centre Ci.
2. Scan frequency components from a lower frequency to a higher frequency,
and let the next unscanned component be Fi. Then, calculate m^j, which
is the distance between Fi and the existing cluster centre Cj. If rriij is
greater than me for all j , then the Fi is chosen as a new cluster centre

3. Repeat Step 2 until all frequency components have been scanned.

4. For each cluster centre, let all frequency components belong to the cluster
if the distance between the components and the cluster centre is less than
me.
312 Kunio Kashino

Here, me is a threshold for sound segregation chosen between 0 and 1. If it is

close to 0, then many clusters are generated. Note that a frequency component
can be shared by multiple clusters.
It is obvious that there are problems and limitations with the above model.
First, the function is only an approximation of human responses in a specific
and simplified experimental setting, and is not guaranteed to be sufficiently
accurate in other, more realistic, situations. Second, with real music signals, it
is diSicult to determine the onset time and onset gradient for each component
because frequency components tend to exhibit complex amplitude patterns.
Third, a sound without a fundamental frequency, such as a percussive sound or
a missing-fundamental sound, will not form a cluster in the above algorithm.
Fourth, other cues such as sequential integration and timbre memories [333]
are not incorporated.

10.2.3 Top-Down Clues for Music Scene Analysis

Bregman pointed out that hearing a signal in a mixture of sounds depends

on conscious eff'ort and prior learning, and called the process behind this
schema-based segregation and integration [49]. From an engineering viewpoint,
the schema-based process often corresponds to knowledge-based or top-down
processing.
Here, we distinguish between the ideas of knowledge-based and top-down
processing. The words bottom-up and top-down signify the direction of the
processing between the levels of abstraction. In bottom-up processing, low-
level features are transformed to a higher level to form larger-scale or more ab-
stract entities, whereas in top-down processing, entities at a high-abstraction
level predict, verify, or control lower-level description. On the other hand,
knowledge-based means that the process uses some prior information. That
is, both bottom-up and top-down processes can be knowledge-based in the
sense that they can use prior information stored in the system. To clarify
this, in this chapter we will use the term 'internal model' to refer to such
prior information.
Many researchers have addressed the integration of bottom-up and top-
down processing modules. Regarding auditory scene analysis tasks, Lesser
et al. built a sophisticated audio signal understanding system based on the
blackboard architecture [407]. The blackboard architecture was employed to
integrate and control the diff'erent processing modules; for example, the system
creates high-level hypotheses which are then used for tuning the front-end
signal processing parameters.
Godsmark et al. proposed a blackboard model for computational auditory
scene analysis [215]. It was designed to accommodate various grouping cues
including pitch proximity, timbral similarity, and source-specific knowledge
such as metre and melodic phrases. The grouping mechanisms interact in a
 10 Auditory Scene Analysis in Music Signals 313

context-sensitive and retroactive^ manner on a blackboard containing eight

levels of abstraction. For example, the model allows high-level predictions
about metre and melody to influence the primitive organization at lower levels
of the blackboard.
Ellis presented a prediction-driven architecture for computational auditory
scene analysis [171]. His motivation included the detection of non-tonal events;
if we use harmonicity to group frequency components as reviewed above, non-
harmonic sounds such as drums, noises, and clicks will not be recognized
as one sound. His idea was to generate both noise component and periodic
component hypotheses and then to look at the acoustic features in order
to either justify or reject the hypotheses. His system was able to simulate
old-plus-new heuristics [49],^ sequential integration of successive tone events
into a stream [49], and auditory restoration phenomenon, which means that
a partially masked sound event tends to be restored based on the elements
before and after the masking event [662]. Recently, the idea has been further
extended based on a quantitative formulation [539].
For bottom-up and top-down integration it is essential to formulate a quan-
titative method that clarifies what is computed or optimized by the algorithm.
In the following sections in this chapter, we review probabilistic approaches
that form a quantitative base for integration. It should be noted, however, that
sometimes the rules can still be effective even in a probabilistic framework.

10.3 Probabilistic Models for Music Scene Analysis

Recently, researchers have proposed probabilistic models for music scene

analysis. This section begins by introducing the Bayesian network, which is
a flexible tool for obtaining probabilistic inferences. Then, its applications to
music scene analysis are reviewed.

10.3.1 Posterior Estimation by the Bayesian Network

A Bayesian network is a directed acyclic graph (DAG) where the nodes cor-
respond to random variables and the links between the nodes encode proba-
bilistic dependences between corresponding random variables [509], [205].
A random variable corresponds to an event to be modelled. A directed link,
represented by an arrow, shows the direction of the probabilistic dependency.
The origin of the arrow is called the parent and the end point is called the child.
Each link can encode conditional probabilities, which are the probabilities of

^In a retroactive system, the interpretation of previous material can be affected

by what happens afterwards.
^Old-plus-new heuristics loosely refers to the principle that, whenever possible, a
change in the signal is interpreted as a continuation of the previously played sounds
plus new sound elements.
314 Kunio Kashino

Fig. 10.7. An example Bayesian network.

child events given the parent events. The word acyclic means that there is no
route from any node that returns to the original node as long as the links are
followed in the designated direction. Such a graph can be singly or multiply
connected. Singly connected means there is only one path between any two
nodes in the graph. Otherwise the graph is referred to as multiply connected.
When a graph is directed and singly connected, it is called a tree if none of
the nodes has more than one parent.
The objective of considering such a network is to calculate posteriors effi-
ciently after some of the random variables have been fixed or observed. The
absence of a link between two nodes means that there is no direct relationship
between the corresponding random variables. Even if these nodes are linked
indirectly (i.e., via other nodes), they are independent when at least one node
existing between the two is fixed. Posterior calculation on the Bayesian net-
work takes advantage of this property.
Figure 10.7 shows an example of a Bayesian network. If the network is
singly connected as in the figure, the posterior calculation is straightforward.
As an example, assume we wish to find the posterior probabilities induced at
node B. Letting D^ represent the data contained in the tree rooted at B and
D^ for the data contained in the rest of the network, we have

P{B) = PiB\D+,Ds). (10.9)

Based on the singly connected structure of the Bayesian network, we assume

the independence of random variables given B, namely

P{D+,D^\B) = PiD+\B)P{D^\B), (10.10)

and then, according to Bayes's theorem, we have

P{B\D+,D^) = aP{D^\B)P{B\D%), (10.11)

where the constant a = 1/P{DQ,D^) can be determined so that the left-hand

side of (10.11) is normalized.
 10 Auditory Scene Analysis in Music Signals 315

Substituting as X{B) = P{D~\B) and 7T{B) = P{B\D^), Equation (10.11)

can be written as
P{B) = aX{B)7T{B). (10.12)
Now we want to obtain X{B) and 7T{B). First, we consider X{B). Denoting
the data encoded in the tree rooted at the k~th child oi B by D^~, we have
X{B) = P{D-\B) (10.13)
= f3Y[P{D^-\B) (10.14)
k

= (3l[Xk{B), (10.15)
k

where /? is a normalization constant that can be chosen so that the sum

2 7 H^j) equals 1. Here bj denote the different values of ^ . In the last equa-
tion, we defined Xk{B) = P{D^~\B). In a singly connected graph, there is
no direct link between children, and therefore given a fixed parent node, the
children are independent. This makes it possible to factor the children. If we
assume E is the A:-th child and e^ are diff"erent values of E, we have

P{D''-\B)=Y^P{D-E\B,ei)P{ei\B) (10.16)
i

=Y,P{D-E\ei)P{ei\B) (10.17)
i

=Y,Kei)P{ei\B), (10.18)
i

and therefore, using conditional probabilities of children given the parent such
as P{ei\hj) and (10.15), we obtain A's node by node. Here, P{ei\hj) is to be
obtained from a statistical model or data, that is learned by or provided to
the system.
Now we consider 7r(5). We have
i^{B)=P{B\D%) (10.19)

=Y,P{B\ai,Dl)P{ai\Dl) (10.20)
i

=Y,P{B\ai) P{ai\D+) (10.21)

i

=Y, P{B\ai) 177r(ai) J ] A(a,) I , (10.22)

i \ rn )

where m is a suSix enumerating the siblings of B except for B, ai are different

values of A, and 7 is a normalization constant. The last equation follows from

P(A|D+) = _ M _ . (10.23)
316 Kunio Kashino

The term in parenthesis in (10.22) is already calculated when P{A) is cal-

culated. This means that TT can also be calculated node by node, using the
conditional probabilities of children given the parent such as P{bj\ai).
Now we have shown that X{X) can be derived from A (children of X) and
7r(X) from 7r(parent of X), using conditional probabilities P(child[parent) be-
tween two adjacent nodes. This allows us to calculate posteriors using (10.12).
The calculation of A(-) and 7r(-) can be viewed as the propagation of diagnostic
and causal support for X, respectively.
In the above discussion, we assumed that the network is singly connected.
If the network is multiply connected, the algorithm presented above will not
properly terminate due to the loops of the probability propagation paths.
Various methods have been developed to obtain the posteriors for such cases,
including the junction tree algorithm and approximative methods using Monte
Carlo sampling and variational methods [396].
The junction tree algorithm utilizes a tree whose nodes are a group of nodes
in the original graph. First, the original graph is converted to an undirected
graph by operations called moralization and triangulation. Then, a junction
tree is constructed by substituting the cliques^ in the converted graph with a
node.
In the junction tree, posterior probabilities can be factored using a func-
tion defined for each node called a clique potential, and calculated with the
probability propagation process designed for undirected trees.

10.3.2 Bayesian Networks Applied to Music Signals

Here, we review how the Bayesian networks were applied to a music scene
analysis task. The first example is a processing model called Organized
Processing Toward Integrated Music Scene Analysis (OPTIMA) [332]. The
input of the model is assumed to be monaural music signals. The output is a
music scene description, that is, a hierarchical representation of musical events
such as frequency components, notes and chords. As shown in Fig. 10.8, the
model consists of three blocks: (A) a pre-processing block, (B) a main process-
ing block, and (C) internal models.
In the pre-processing block, frequency analysis is performed and a sound
spectrogram is obtained. Then, frequency components are extracted. An ex-
ample of the power transition of a frequency component is shown in Fig. 10.9.
With complicated spectrum patterns, it is difficult to recognize the onset and
offset times solely based on bottom-up information. Thus the system creates
several terminal point candidates for each extracted component.
Rosenthal's rhythm recognition method [547] and Desain's time quanti-
zation method [140] are used to obtain rhythm information for the precise
extraction of frequency components and recognition of the onset/offset time.

^A clique of a graph is its maximal complete subgraph. A complete graph is a

graph in which each pair of nodes is connected by a link.
 10 Auditory Scene Analysis in Music Signals 317

Main Processes Internal Models

Temporal Processing Modules

Chord Transition
Prediction
] C Chord Group
Creation

Bottom-Up Hypothesis Network Top-Down

Processing Modules Processing Modules

Chords

Chord Analysis
mS2L
Musical Notes ^
Note Prediction

Clustering for
Source Identification
Clustering for
IXIXIXI Freq.Component
Prediction
Tone Memories

Sound Formation
Freq. Components

Freq.Component
Extraction
Processing Window
Creator
K Rhythm / Beat
Analysis
J
Time-Freq. Analysis
0 Output Data
Generation
) (
Preprocesses

7^
Monaural Music Signals
loWwimramW
Resynthesized
Sound
M\D\ Data Display

Fig. 10.8. An overview of the OPTIMA processing model.

i
^ 1
o Frequency component

Lo ^o o A / W V A A
(r o o Q 0 O O P
L
0 1 2 3 Time[s]
O Termination candidate point

Fig. 10.9. A frequency component and its terminal point candidates.

Based on the integration of the beat probabilities and termination probabil-

ities of terminal point candidates, the candidates are determined as being
continuous or terminated, and consequently processing windows are formed.
Here a processing window is a group of frequency components with similar
318 Kunio Kashino
processing windows

0 1 2 3 4 5 Time[s]
Fig. 10.10. A spectrogram of a polyphonic music excerpt and the processing win-
dows.

onset times. The processing window is utilized as a time base for the subse-
quent main processes.
When each processing window is created in the pre-processing block, it
is passed to the main processing block. The main block involves a Bayesian
network. Exactly as discussed in the previous section, the Bayesian network
has three layers: component, note, and chord levels. The chord level nodes
are connected in time as the time proceeds. Each node in the network is a
random variable that encodes multiple hypotheses. That is, the model holds
hypotheses of the external acoustic events as a probability distribution in a
hierarchical space.
The Bayesian network is actually built by multiple processing modules.
The modules are classified into two types: those for creating the nodes and
providing initial probabilities to the nodes, and those for providing condi-
tional probabilities to the links. The former is called creators, and the latter
predictors.
There are two creators in OPTIMA: a note hypothesis creator and a chord
hypothesis creator. As described above, first, frequency component hypothe-
ses and processing windows are created. Then the note hypothesis creator
generates the hypotheses by referring to perceptual rules such as harmonic
mistuning and onset asynchrony as described in the previous section. The
creator also consults timbre models for a timbre discrimination analysis to
identify the sound source of each note. A chord hypothesis creator generates
the chord hypotheses when note hypotheses are given. This creator refers to
chord naming rules.
 10 Auditory Scene Analysis in Music Signals 319

Table 10.3. Examples of the chord-note relation knowledge. The conditional prob-
abilities P(note|chord) were obtained by statistical analysis of printed music.

Given chord (5 examples)

Note A A7 Am Am7 Adim
A .983 di .023 .852 .074 1.00 .000 .933 lb .081 .781 lb .143
A# .000 .000 .023 =b .031 .030 .058 .000 .000 .188 d= .135
B .150 .064 .364 .101 .091 .098 .182 .132 .094 .101
C .008 .016 .023 .031 .848 zb .122 1.00 .000 .656 .165
c# .850 .064 .818 .081 .000 .000 .030 lb .058 .031 lb .060
D .025 .028 .057 .048 .182 .132 .394 .167 .031 lb .060
D# .067 lb .045 .023 .031 .030 .058 .000 .000 .406 lb .170
E .842 .065 .545 .104 .909 .098 .879 .111 .031 .060
F .017 .023 .023 .031 .000 .000 .030 .058 .094 .101
F# .125 .059 .148 =b .074 .000 .000 .061 .081 .469 .173
G .025 lb .028 .773 .088 .121 .111 .939 .081 .062 .084
G# .075 .047 .045 lb .044 .000 lb .000 .000 =b .000 .156 .126
: 95% reliable range

There are three predictors in OPTIMA that provide the conditional

probabilities of a child, given a parent, to the links. A frequency com-
ponent predictor calculates P(component|note). A note predictor evaluates
P(note|chord). A chord transition predictor provides chord transition prob-
abilities P(chord|chord). These processing modules use six types of internal
models.
The chord transition model holds statistical information on chord pro-
gressions, under the tri-gram assumption (see Chapter 5 for further details
on A^-gram models). This dictionary is based on a statistical analysis of 206
traditional songs (all Western tonal music). It is used by a chord transition
predictor.
The chord-note relation model is used by a note predictor. The model
involves a database that stores the probabilities of notes that can be played
in a given chord. This information is also obtained by a statistical analysis of
the 206 songs. Part of the stored data is shown in Table 10.3.
The chord naming rules are based on music theory, and are used by a
chord hypothesis creator to recognize chords when the hypotheses of played
notes are given.
The tone memory stores instances of frequency component data of a single
note played by various musical instruments such as a clarinet, flute, piano,
trumpet, or violin at different degrees of loudness (forte, medium, piano),
frequency range, and durations. This memory is used by a frequency compo-
nent predictor.
The timbre models are formed in the feature space of the timbre. An eleven-
dimensional feature space was created using principal component analysis, and
320 Kunio Kashino
time

chord

note

component

Q : node ^ : link

Fig. 10.11. An example of created Bayesian networks.

each of the above-mentioned musical instruments was modelled by a Gaussian

distribution in the feature space. This model is used by a note hypothesis
creator to provide the initial probabilities of sound sources (i.e., instrument
names).
Finally, the perceptual rules are used by a note hypothesis creator to create
note hypotheses from the frequency component hypotheses. The rules involve
the clustering algorithm described in the previous section [334].
Each time a frequency component hypothesis is created in a processing
window, the hypothesis creators create the node instances and links. When a
link is created, the predictors provide conditional probabilities. Then, a prob-
ability propagation series is automatically performed and posterior probabil-
ities at that time are calculated. As shown in Fig. 10.11, multiple successive
note-level nodes can be linked to a single chord level node based on rhythm in-
formation extracted during the pre-processing stage. Thus, the instance of the
network structure grows dynamically as the input signal arrives. Although the
higher-level hypotheses (e.g. chords) are created based on lower-level informa-
tion (e.g. notes), the higher-level information is still useful for the lower-level
hypotheses, because it reflects and integrates a larger-scale context and infor-
mation from the corresponding internal models.
In the OPTIMA model, temporal transitions are solely represented at the
chord level. This simplifies the network structure to a tree, but generally this
is insufficient because note-level temporal dependency is not considered.
The second example we review here is a music stream network [331], in
which note-level temporal transitions were introduced. In this example, the
Bayesian network was employed for sound source identification for ensemble
music.''

^The work toward a musical sound source identification is discussed in Chapter 6

in this book.
 10 Auditory Scene Analysis in Music Signals 321

Consider two musical notes n^, Uk-i {k denotes the order of the onset
times of these notes, Uk-i preceding n^). An 'impedance' measure Z(n/e, n^-i)
is defined as

Z{nk,nk-i)^wY,{- ^i logPi(nfc,nfc_i)} , (10.24)

i

where i is a suffix that enumerates diflferent terms of Z, Pi is the conditional

probability of the occurrence of the transition from Uk-i to Uk in a given
musical context, and Wi {> 0) is the weight for each term. Z can be viewed
as a weighted sum of the amount of information, logP^, delivered by the
transition from rik-i to n^. Thus, Z reflects the infrequency of the transition
for these two notes. Then, a 'music stream' [447] is formed as a sequence of
musical notes that gives a local minimum of Z.
The term VF is a time window that is defined as

W{6t)=exp(-^ , (10.25)

where St is the difference between the onset times of these two notes, and r
is a time constant. Unlike ordinary time windows, W becomes greater as 5t
increases. This loosely corresponds to the proximity rule of auditory stream
organization as described in [49].
In this example, the following three Z factors are considered: (Pi) the
transition probabilities of musical intervals, (P2) the transition probabilities
of timbres, and (P3) the transition probabilities of musical roles.
The first factor is the musical interval probability. In tonal music, the
musical intervals of note transitions do not appear equally often; some inter-
vals are more frequent than others. Thus the pitch transition probability in a
melody can be utihzed as Pi in (10.24). The probabilities Pi were obtained
from 397 melodies extracted from 196 pop scores and 201 jazz scores, where
the total number of note transitions was 62,689. Figure 10.12 shows the es-
timated probabilities. The analysis was made only for the principal melodies
and may not be precisely valid for the other melodies such as bass lines or
parts arranged for polyphonic instruments such as the piano. For simplicity,
however, probabilities shown in Fig. 10.12 for Pi were used for all cases.
The second factor is the timbre transition probability. It is reasonable to
suppose that a sequence of notes tends to be composed of notes that have
similar timbres. To incorporate this tendency, a distance measure was defined
between the timbres of two notes, so as to estimate the probability that two
notes a certain distance apart would appear sequentially in a musical stream.
These probabilities form P2 in (10.24).
The distance between timbres is defined as the Euclidean distance between
the timbre vectors in a timbre space. A timbre space can be spanned in several
ways. In the experiment described in [331], each axis of the space corresponds
to a musical instrument name, and a timbre vector is composed of correlation
322 Kunio Kashino
Transition
Probability [%]
20

14 12 10 4 6 8 10 12 14
H + Interval (semitone)

Fig. 10.12. Probabilities of musical intervals.

Probability

Distance

Fig. 10.13. Conversion from distance to probability. The calculated distance d

is converted to Pg, which is the probability of the appearance of the distance in
a sequence of musical notes, by using a histogram of distances. The histogram is
normalized so that the histogram values add up to one.

values between the input signal and each of the template signals of musical
instruments stored in advance. Then, the distances between the timbre vectors
of successive notes in a sequence are translated into probabilities using a
normalized histogram as explained in Fig. 10.13. This histogram models the
distribution of the timbre vectors for successive notes.
The third factor is musical role consistency. In ensemble music, a sequence
of notes can be regarded as carrying a musical role such as a principal melody
or a bass line. To introduce such musical semantics, the probability P3 is
introduced:
Ps = ar-^b, (10.26)
where a and b are constants, and r is the rate of the highest (or lowest) notes in
the music stream under consideration. Equation (10.26) represents a musical
heuristic that the music stream formed by the highest (lowest) notes tends to
continue to flow to the highest (lowest) note.
 10 Auditory Scene Analysis in Music Signals 323

> l Step 2
\-3 a "k
-^^ .^?^^
^ ;
n\"^ ^
vfl5 o_ %V^^^.
Step 3

Fig. 10.14. A procedure for creating music stream networks. See text for details.

Using (10.24), the networks that correspond to sequences of musical notes

are built by a procedure that is illustrated in Fig. 10.14:
1. When a new node Uk has just been created, the system first chooses
the link that gives the minimum Z value (/i) among the candidate links

2. The system then evaluates Z values for the link candidates {gir'' ^ds)
from the selected node {uks), to choose the hnk with the minimum Z
value (^i).
3. If ^1 and /i are identical, the link composes a music stream. If a music
stream from riks has already been formed in a direction other than pi,
the stream is cut; the direction of the music stream is changed to gi{= h).
Thus the networks are built by connecting nodes that give the locally mini-
mum Z value.
Once the network has been built, then it can be considered as a Bayesian
network and the posterior probabilities of sound sources are calculated. An
example of the system in operation is shown in Figure 10.15. The input here
is a monaural recording of a real ensemble performance of 'Auld Lang Syne',
a Scottish folk song, arranged in thee parts and performed by a violin, a flute,
and a piano. Figure 10.15 displays the recognized music streams as well as
the status of nodes for the beginning part of the song. The bars in each node
indicate the probabilities at the node (not normalized). The links between the
nodes are the extracted music streams. It is shown that each part is correctly
recognized as the music stream. The thickness of the link line corresponds to
its Z value given by (10.24); a thick line represents a link with a low Z value.
324 Kunio Kashino
I H

Fkitei
PfanoT I

/ \ _A /
1-S5| \ I -SS J
g

I
-S 1 1 10-53

^ A 1
rt:^
Piano

Fig. 10.15. An example of a created music stream network.

10.4 Conclusion: From Grouping to Generative

Estimation
As shown throughout this chapter, early work on computational auditory
scene analysis was rather directly inspired by psychological and physiological
findings. In particular, frequency component grouping rules were intensively
investigated. Various systems have been designed to simulate simultaneous,
sequential, and schema-based grouping and integration.
Fortunately, probabilistic approaches have become computationally feasi-
ble in recent years [101], [672], [208]. For example, as discussed in Chapter 7,
the Bayesian approach has been successfully applied to fundamental frequency
estimation for music signals [212]. Although the first trials of the Bayesian ap-
proach for music scene analysis were simplified in terms of model structure
and parameter exploration [332], the approach has been greatly extended in
recent years. For example, Sterian developed a music transcription system
based on a probabilistic framework [611]. The system tracks frequency com-
ponents with a Kalman filter, which is equivalent to Bayesian estimation, and
then it maintains multiple hypotheses for note formation. Cemgil proposed a
generative model for music transcription [74], as described in Chapter 7.
 10 Auditory Scene Analysis in Music Signals 325

We anticipate that it will be possible to explain certain auditory restora-

tion phenomena in terms of Bayesian inference. When this is proven to be
the case, the probabilistic inference approach to music scene analysis may be
viewed as a newer version of the ASA-inspired approach.
11
Music Scene Description

Masataka Goto

National Institute of Advanced Industrial Science and Technology (AIST).

1-1-1 Umezono, Tsukuba, Ibaraki 305-8568, Japan. m.gotoQaist.go. jp

11.1 Introduction

This chapter introduces a research approach called 'music scene description^

[232], [225], [228], where the goal is to build a computer system that can under-
stand musical audio signals at the level of untrained human listeners without
trying to extract every musical note from music. People listening to music can
easily hum the melody, clap hands in time to the musical beat, notice a phrase
being repeated, and find chorus sections. The brain mechanisms underlying
these abilities, however, are not yet well understood. In addition, it has been
difficult to implement these abilities on a computer system, although a system
with them is useful in various applications such as music information retrieval,
music production/editing, and music interfaces. It is therefore an important
challenge to build a music scene description system that can understand com-
plex real-world music signals like those recorded on commercially distributed
compact discs (CDs).
Music scene description differs from two popular approaches to deal with
music signals, sound source separation and traditional automatic music tran-
scription (in the narrow sense^). Although these technologies are valuable from
an engineering viewpoint, neither separation nor transcription is necessary or
sufficient for understanding music.
It is possible to understand music without sound source separation.
The fact that human listeners understand various properties of audio sig-
nals is not necessarily evidence that the human auditory system extracts
the audio signal of each individual source. Even if a mixture of two compo-
nents cannot be separated, it can be understood from their salient features
that the mixture includes them. In fact, from the viewpoint of auditory
^The term 'automatic music transcription' in this chapter refers to a traditional
approach of transcribing all musical notes as a score, while the term 'automatic
music transcription' in this book has a broader meaning including the music scene
description as described in Chapter 1 of this volume.
328 Masataka Goto

Compact disc (CD) 1 (Melody line ^-^'^f ) 1 Music scene descriptions

1 [Bass line _-__-A ) |
same same
Repeated
1 Hierarchical - ' 11 sections
I beat structure J i ' ' i ' i i i J1

Drums " ' ." .' . ' 1

Musical
audio signals Local descriptions 1 Global descriptions

Fig. 11.1. Music scene descriptions.

psychology, it has been pointed out that human hsteners do not perform
sound source separation: perceptual sound source segregation is differ-
ent from signal-level separation. For example, Bregman noted that 'there
is evidence that the human brain does not completely separate sounds'
[50]. The approach of developing methods for monaural or binaural sound
source separation might deal with a hard problem which is not solved by
any mechanism in this world (not solved even by the human brain).
It is possible to understand music without complete music transcription.
Music transcription, identifying the names (symbols) of musical notes, is a
difficult skill mastered only by trained musicians. As pointed out by Goto
[239], [240], [232] and Scheirer [565], untrained listeners understand music
to some extent without mentally representing audio signals as musical
scores. For example, as known from the observation that a listener who
cannot identify the name and constituent notes of a chord can nevertheless
feel the harmony and chord changes, a chord is perceived as combined
whole sounds (tone colour) without reducing it to its constituent notes
(like reductionism). Furthermore, even if it is possible to derive separated
signals and musical notes, it would still be difficult to obtain high-level
music descriptions like melody lines and chorus sections.
The music scene description approach therefore emphasizes methods that
can obtain a certain description of a music scene from sound mixtures of
various musical instruments in a musical piece. Here, it is important to discuss
what constitutes an appropriate description of music signals. Since various
levels of abstraction for the description are possible, it is necessary to consider
which level is an appropriate first step towards the ultimate description in
human brains. Goto [232], [228] proposed the following three viewpoints:
An intuitive description that can be easily obtained by untrained listeners.
A basic description that trained musicians can use as a basis for higher-
level music understanding.
A useful description facilitating the development of various practical ap-
plications.
 11 Music Scene Description 329

According to these viewpoints, the following local and global descriptions

(Fig. 11.1) have been proposed for Western music:
1. Melody and bass lines
Melody and bass hnes represent the temporal trajectory of the melody and
bass. The melody is a series of single tones and is heard more distinctly
than the rest. The bass is a series of single tones and is the lowest frequency
part in polyphonic music. Note that a melody or bass line here is not
represented as a series of musical notes; it is a continuous representation
of fundamental frequency (FO, perceived as pitch) and power transitions.
Only music with distinct melody and bass lines is dealt with for this
description.
2. Hierarchical beat structure
Hierarchical beat structure represents the fundamental temporal structure
of music and comprises the quarter-note (beat) and measure levelsi.e.,
the positions of quarter-note beats and bar lines (corresponding to the
metrical levels of 'beat' and 'bar' in Fig. 4.1, p. 106).
3. Drums
Drums represent onset times of principal drum sounds, such as bass and
snare drums. Their temporal patterns form drum patterns. Only music
with drum sounds is dealt with for this description.
4. Chorus sections and repeated sections
Chorus sections represent the most representative, uplifting, and promi-
nent thematic sections in the structure of a musical piece (especially in
popular music). Since chorus sections are usually repeated, they are rep-
resented as a list of the start and end points of every chorus section.
Repeated sections represent the repetition of temporal regions with vari-
ous lengths. Only music with distinct repeated choruses, such as popular
music, is dealt with for the description of chorus sections, while any music
can be dealt with for the description of repeated sections.
The idea behind these descriptions came from introspective observation of how
untrained listeners listen to music. The following sections introduce methods
for producing these descriptions from music signals such as CD recordings,
which contain simultaneous sounds of various instruments (with or without
drum sounds). In general, these methods deal with monaural audio signals
because stereo signals on CDs can be easily converted to monaural signals by
averaging the left and right channels. While methods depending on stereo
information [24] can have better performance than methods dealing with
monaural signals, such stereo-based methods cannot be applied to monau-
ral signals. Methods assuming monaural signals, on the other hand, can be
applied to stereo signals and be considered essential to music understanding
since human listeners have no difficulty understanding the above descriptions
even from monaural signals.
330 Masataka Goto

11.2 Estimating Melody and Bass Lines

The estimation of melody and bass lines is important because the melody
forms the core of Western music and is very influential in the identity of a
musical piece, while the bass is closely related to the tonality (see Chapter 1).
These lines are fundamental to the perception of music by both musically
trained and untrained listeners. They are also useful in various apphcations
such as automatic music indexing for information retrieval (e.g., search-
ing for a song by singing a melody), computer participation in live human
performances, musical performance analysis of outstanding recorded perfor-
mances, and automatic production of accompaniment tracks for karaoke using
CDs.
It is difficult to estimate the fundamental frequency (FO) of melody and
bass lines in monaural sound mixtures from CD recordings. Most previ-
ous FO estimation methods cannot be applied to this estimation because
they require that the input audio signal contain just a single-pitch sound
with aperiodic noise or that the number of simultaneous sounds be known
beforehand. The main reason FO estimation in sound mixtures is difficult
is that, in the time-frequency domain, the frequency components of one
sound often overlap the frequency components of simultaneous sounds. In
popular music, for example, part of the voice's harmonic structure is often
overlapped by harmonics (overtone partials) of the keyboard instrument
or guitar, by higher harmonics of the bass guitar, and by noisy inhar-
monic frequency components of the snare drum. A simple method for lo-
cally tracing a frequency component is therefore neither reliable nor stable.
Moreover, FO estimation methods relying on the existence of the FOs fre-
quency component (the frequency component corresponding to the FO) not
only cannot handle the missing fundamental, but are also unreliable when
the FOs frequency component is smeared by the harmonics of simultaneous
sounds.
FO estimation of melody and bass lines in CD recordings was first achieved
in 1999 by Goto [232], [222], [228]. Goto proposed a real-time method called
PreFEst (Predominant-FO Estimation method) which estimates the melody
and bass lines in monaural sound mixtures. Unlike previous FO estimation
methods, PreFEst does not assume the number of sound sources, locally trace
frequency components, or even rely on the existence of the FOs frequency
component. PreFEst basically estimates the FO of the most predominant har-
monic structurethe most predominant FO corresponding to the melody or
bass linewithin an intentionally limited frequency range of the input mix-
ture. It simultaneously takes into consideration all possibilities for the FO and
treats the input mixture as if it contained all possible harmonic structures
with different weights (amplitudes). To enable the application of statistical
methods, the input frequency components are represented as a probability
density function (pdf), called an observed pdf. The point is that the method
regards the observed pdf as a weighted mixture of harmonic-structure tone
 11 Music Scene Description 331

I Audio signals

[ Instantaneous frequency calculation ]

[ Extracting frequency components ]

3 E
Limiting frequency regions
BPF for melody line] ( BPF for bass line

I Frequency components

[Forming FO's probability density function^

Fig. 11.2. Overview of PreFEst (Predominant-FO Estimation method) for estimat-

ing melody and bass lines in CD recordings. In this figure, BPF denotes bandpass
filtering.

models (represented by pdfs) of all possible FOs. It simultaneously estimates

both their weights corresponding to the relative dominance of every possible
harmonic structure and the shape of the tone models by maximum a posteri-
ori probability (MAP) estimation (see Chapter 2, p. 40 for an introduction to
MAP estimation methods) considering their prior distribution. It then consid-
ers the maximum-weight model as the most predominant harmonic structure
and obtains its FO. The method also considers the FOs temporal continuity
by using a multiple-agent architecture.
The following sections first explain the PreFEst method in detail and then
introduce other methods for estimating the melody line developed by Paiva,
Mendes, and Cardoso [494], [493], Marolt [435], [436], and Eggink and Brown
[169], and a method for estimating the bass line developed by Hainsworth
and Macleod [264]. Figure 11.2 shows an overview of PreFEst. PreFEst con-
sists of three components, the PreFEst front end for frequency analysis, the
PreFEst core to estimate the predominant FO, and the PreFEst back end to
evaluate the temporal continuity of the FO. Since the melody line tends to
have the most predominant harmonic structure in middle and high-frequency
regions, and the bass line tends to have the most predominant harmonic struc-
ture in a low-frequency region, the FOs of the melody and bass lines can be
estimated by applying the PreFEst core with appropriate frequency-range
limitation.
332 Masataka Goto

11.2.1 PreFEst Front End: Forming the Observed Probability

Density Functions
The PreFEst front end first uses a multirate filterbank to obtain adequate
time-frequency resolution under a real-time constraint. By using an instanta-
neous frequency-related measure [84], [7], [338] for the existence of frequency
components, it then extracts frequency components ^^^^u) from the short-
time Fourier transform (STFT) X{iy,t) of a signal
T/it)( \ _ / i^(^'OI if ^ ^^^ ^ frequency component, m i^
^ ^" \ 0 otherwise, ^ ' '
where t is the time measured in units of frame-shifts (10 ms), and i/ is the log-
scale frequency denoted in units of cents (a musical-interval measurement).
Frequency /HZ in Hertz is converted to frequency /cent in cents so that there
are 100 cents to a tempered semitone and 1200 to an octave:

/cent = 1200 l0g2 ^ ^ / " % , . (11.2)

440 X 2i2"-^
To obtain two sets of bandpass-filtered frequency components, one for the
melody fine (261.6-4186 Hz) and the other for the bass line (32.7-261.6 Hz)
[228],^ the PreFEst front end uses bandpass filters (BPFs) whose frequency
response is BPF^i(z/) where u denotes the melody line {u = 'melody') or the
bass line {u = 'bass line'). Each set of the bandpass-filtered components is
finally represented as an observed pdf p^ (^)
(.) BPF(.) !^W(.)

11.2.2 PreFEst Core: Estimating the FOs Probability Density

Function
For each melody or bass line set of filtered frequency components represented
as an observed pdf p ^ (^), the PreFEst core forms a probability density func-
tion of the FO, called the FOs pdf, p)pQ{iyo), where I/Q is the log-scale funda-
mental frequency in cents. The PreFEst core considers each observed pdf to
have been generated from a weighted-mixture model of the tone models of all
possible FOs; the tone model is the pdf corresponding to a typical harmonic
structure and indicates where the harmonics (overtone partials) of the FO
tend to occur (Fig. 11.3). Because the weights of tone models represent the
relative dominance of every possible harmonic structure, these weights can be
regarded as the FOs pdf: the more dominant a tone model is in the mixture,
the higher the probability of the FO of its model.
^The method finds the FO whose harmonics are most predominant in those lim-
ited frequency ranges. In other words, whether the FO is within each limited range or
not, PreFEst tries to estimate the FO which is supported by predominant harmonic
frequency components within that range.
 11 Music Scene Description 333

I^one Model (/=2) -p(v,llVo,2,^(Vo,2))

IP(V

Tone Mcxlel 0=1)

p(vlVo,l,|I^'\Vo,

fundamental
frequency
Vn+1',__

Vo+240(r^ V [cent]

Fig. 11.3. Model parameters of multiple adaptive tone models p(i^|i/o, i, M (^o, i))-
Weighted-Mixture Model of Adaptive Tone Models
To deal with diversity of the harmonic structure, the PreFEst core can use
several types of harmonic-structure tone models. The pdf of the i-th tone
model for each FO UQ is denoted by p(i/|i/o,^,M^*H^o,0) (^^^ ^ig. 11.3), where
the model parameter ^^^\VQ^I) represents the shape of the tone model. The
number of tone models is I^^ (that is, i = l , . . . , I i t ) , where u denotes the
melody line {u = 'melody') or the bass line (u = 'bass line'). Each tone model
is defined by
Mu

p(i/|z/o,^,M^*H^o,^)) = 5 Z P(^.^ko,^,M^*H^o,^)), (11.4)

7n=l

p(z/, m|z/o, h M^*^(^0,0) = c^*^{rn\^o^ 0 ^ ( ^ 5 ^o + 1200 log2 m, al), (11.5)

li^'\v^,i) = {c^'\m\uo4) I m-l,...,M,}, (11.6)

where M^ is the number of harmonics considered, cr^ is the variance of the
Gaussian distribution A/'(i/; Z/Q^^U) (^^^ (2.16), p. 29 for a definition), and
c^^\m\vQ^i) determines the relative amplitude of the m-th harmonic compo-
nent (the shape of the tone model) and satisfies

^c(*)(m|i/o,i) = l. (11-7)
771=1

In short, this tone model places a weighted Gaussian distribution at the po-
sition of each harmonic component.
The PreFEst core then considers the observed pdf Px^ {y) to have been
generated from the following model p(i/|0^*^), which is a weighted mixture of
all possible tone models p[v\yQ,i^ii^^\vo^i)):
ph I^

p(H0<*)) = I " Vw;(i/o,i) p{v\v^,i,n^'\uo,i)) di^o, (11-8)

6>(*'= {wW,M}, (11.9)

334 Masataka Goto

i(;W = {w^'Hiyo,i) I F^, < i^o < F^,f = 1 , . . . ,I}, (11.10)

M ^ = {/xW(j/o,i) \F',<uo<K,i = 1,...,I}, (11.11)

where F\^ and FJ^ denote the lower and upper Hmits of the possible (allowable)
FO range and w^^\iyo, i) is the weight of a tone model p(i/|z^o, h I^^^H^o^ ^)) that
satisfies

L ^'Tw^'Huo.i)diyo = l- (11.12)

Because the number of sound sources cannot be known a priori, it is important

to simultaneously take into consideration all FO possibilities as expressed in
(11.8). If it is possible to estimate the model parameter 6^^^ such that the
observed pdf p^ (ly) is likely to have been generated from the model p(i/|0^*^),
the weight w^*\uo,i) can be interpreted as the FOs pdf pi?o(^o)-

PFO(^O) = J^w^'\iyo.i) (K < ^0 < F^). (11.13)

Introducing a Prior Distribution

To use prior knowledge about FO estimates and the tone model shapes, a prior
distribution pou(^ ) of ^^^^ is defined as follows:

Po(0^*') = \>ou{w^'^) Po(At('>), (11.14)

Here, poui'w^^^) and pouifi^^^) are unimodal distributions: poit(ty^^^) takes

its maximum value at K;QJ(Z/O,0 ^^^ POU(M^*^) takes its maximum value
a^ MoLVo,^ (= {col{m\iyo,i) \ m = 1,...,M^}), where i(;^*J(i/o,i) and
MouC^o^O ^^^ ^he most probable parameters. Figure 11.4 shows two exam-
ples of the most probable tone model shape parameters, fjLQ^{i/o,i), used in
Goto's implementation. Z^ and Z^ are normalization factors, and fi^u and
Pliuii^o, i) are parameters determining how much emphasis is put on the max-
imum value. The prior distribution is not informative (i.e., it is uniform) when
Piuu and PfiuiJ^Oi 0 are 0, corresponding to the case when no prior knowledge
is available. In practice, however, /?^w(i^o?0 should not be 0 and a prior dis-
tribution of the tone model shapes should be provided. This is because if the
prior distribution of the tone model shapes is not used, there are too many
 11 Music Scene Description 335

.1 ,1 illIhA,...
LJ ij
frequency [cent]
1 1,1 IIIIIAU...,
I1
frequency [cent]

(a) for melody line (b) for melody line

{u = 'melody', i = 1, Z/Q = 4000 cent) {u = 'melody', z = 2, I/Q = 4000 cent)

2000
l.l
2500
I
3000 3500 4000 4500
1,l,,
frequency [cent]
5000 5500 6000 2000 2500 3000 3500 4000 4500
frequency [cent]
A
5000
,
5500 6000

(c) for bass line (d) for bass line

(u = 'bass line', i = 1, I/Q = 2000 {u = 'bass line', i = 2,1/0 = 2000
cent) cent)

Fig. 11.4. Examples of prior distribution of the tone model shapes

degrees of freedom in their shapes. W i t h o u t the prior distribution, unrealistic

tone model shapes, such as a shape having only one salient component at
frequency of the fourth harmonic component, could be estimated. In (11.15)
and (11.16), D^{W''QI',W^^^) and D^{fi^Q^{uo,i)',ii^^\i/o,i)) are the following
KuUback-Leibler information:

nW/ duo (11.17)

Mu
Col{'m\vo,i)
D^(Mo('^o,i);M^*'K,i)) = E 4 ( " ^ k o , i ) log (11.18)
m=l c(*)(m|i/Oii)

These prior distributions were originally introduced for the sake of analyt-
ical tractability of the expectation maximization (EM) algorithm to obtain
intuitive (11.25) and (11.26).

M A P Estimation Using the E M Algorithm

T h e problem to be solved is to estimate the model parameter 6^^\ taking into

account the prior distribution po^i'^ )^ when \y^\u) is observed. T h e M A P
estimator of 6^^^ is obtained by maximizing
336 Masataka Goto

p<*V) (logp(i^|e^*') + logpo(0(*')) du. (11.19)

/
Because this maximization problem is too difficult to solve analytically, the
PreFEst core uses the expectation maximization (EM) algorithm (see the
presentation of the EM algorithm in Chapter 2, p. 35 and [138]), which is an
algorithm where two stepsthe expectation step (E-step) and the maximiza-
tion step (M-step)are iteratively applied to compute MAP estimates from
incomplete observed data (i.e., from p^(i/)). With respect to 0^*\ each itera-
tion updates the old estimate 6'^^^ {w'^^\ fi'^^^ to obtain a new (improved)
estimate 0 = {w^^\Jl^ ^}. For each frame t, it;'^*^ is initialized with the final
estimate {t;^*~^^ after iterations at the previous frame t 1; ^'^^^ is initialized
with the most probable parameter /XQJ in the current implementation.
By introducing the hidden (unobservable) variables Z/Q, ^ and m, which,
respectively, describe which FO, which tone model, and which harmonic com-
ponent were responsible for generating each observed frequency component at
v^ the two steps can be specified as follows:
1. E-step:
Compute the following (5MAP(^^^^|0'^^^) for the MAP estimation:

QMAP(^^*V'^'^) - Q{e^'^\e'^'^)^\ogpou{e^'^). (11.20)

oo

/ -00

(11.21)
where Q{6^^^ \0'^^^) is the conditional expectation of the mean log-likelihood
for the maximum likehhood estimation. Ei^Q,i,rn[<^|^] denotes the condi-
tional expectation of a with respect to the hidden variables Z/Q, h ^^^ ^?
with the probability distribution determined by condition b.
2. M-step:
Maximize QMAP(^ |^ ) as a function of 6^*^ to obtain an updated (im-
proved) estimate 9 :

e^'^ = argmax QMAP(^^*^|0'^*^). (11.22)

In the E-step, Q(0^*V^*^) is expressed as

oo ft ^u ^y^u

p(z/o, ^, m|z/, ^'(*^) log p(i/, Z/Q, ^, m|^^*^) duodu, (11.23)

 11 Music Scene Description 337

where the complete-data log-hkeUhood is given by

logp(i/,z/o,i,m|0^*^) =log{w^^\iyo,i) p{iy,m\uo,i, fi^^\uo,i))). (11.24)

Regarding the M-step, (11.22) is a conditional problem of variation, where

the conditions are given by (11.7) and (11.12). This problem can be solved by
using Euler-Lagrange differential equations with Lagrange multipliers [222],
[228] and the following new parameter estimates are obtained:

1 1 PlVU

where Wul{uo^i) and c i i ^ ^ l ^ o , ^ are, when the noninformative prior distrib-

ution {pwu = 0 and Pllu{i^o, i) = 0) is given, the following maximum likelihood
estimates:

WIL{UQ,1)= / py{jy)^ \ / ~ du,

J~^ II: E L I ^'^'^ iv. k) p(^|7/. A:, //'(*) (ry, k)) dv
(11.27)
-(*)/ I ^ ^

f"^ it)., w'^*^{iyo,i) p(z/,m|t/o,i,/x^^^)(^o,i)) ^^

(11.28)

After the above iterative computation of (11.25) and (11.26),^ the FOs pdf
PFO(^O) ^^^ ^^ obtained from w^^\uo^ i) according to (11.13). The tone model
shape c^^\m\iyo^i), which is the relative amplitude of each harmonic compo-
nent of all types of tone models p{iy\iyo,i, fi^^\uo, i)), can also be obtained.

11.2.3 PreFEst Back End: Sequential FO Tracking by

Multiple-Agent Architecture

A simple way to identify the most predominant FO is to find the frequency

that maximizes the FOs pdf. This result is not always stable, however, because
peaks corresponding to the FOs of simultaneous sounds sometimes compete in
the FOs pdf for a moment and are transiently selected, one after another, as
the maximum.
^In implementing the PreFEst core, this iterative computation is simple enough
to perform only (11.25), (11.26), (11.27), and (11.28).
338 Maisataka Goto

7200 cent
melody line g

3600 cent

bass line.

Ocent
^ time [sec] ^
(a) Frequency components (b) Estimated melody and bass lines
(observed pdf p^^ (u) before (the most dominant and stable
applying bandpass filters) FO trajectory in each p}^o(^o))

(c) FOs pdf (p^o(^o)) for estimating (d) FOs pdf (pi^o(^o)) ^^^ estimating the
the melody line in (b) bass line in (b)

Fig. 11.5. Audio-synchronized real-time graphics output for a popular music ex-
cerpt with drum sounds: (a) frequency components, (b) the corresponding melody
and bass lines estimated (final output), (c) the corresponding FOs pdf obtained when
estimating the melody line, and (d) the corresponding FOs pdf obtained when es-
timating the bass line. These interlocking windows have the same vertical axis of
log-scale frequency.

T h e P r e F E s t back end therefore considers the global temporal continuity of

the FO by using a multiple-agent architecture in which agents track different
temporal trajectories of the FO [228]. Each agent starts tracking from each
salient peak in the FOs pdf, keeps tracking as long as it is temporally continued,
and stops tracking when its next peak cannot be found for a while. T h e final
FO o u t p u t is determined on the basis of the most dominant and stable FO
trajectory. Figure 11.5 shows an example of the final o u t p u t .
 11 Music Scene Description 339

11.2.4 O t h e r M e t h o d s

While the PreFEst method resulted from pioneering research regarding melody
and bass estimation and weighted-mixture modelling for FO estimation, many
issues still need to be resolved. For example, if an application requires MIDI-
level note sequences of the melody line, the FO trajectory should be segmented
and organized into notes. Note that the PreFEst method does not deal with
the problem of detecting the absence of melody and bass lines: it simply out-
puts the predominant FO for every frame. In addition, since the melody and
bass lines are generated from a process that is statistically biased rather than
randomi.e., their transitions are musically appropriate this bias can also be
incorporated into their estimation. This section introduces other recent ap-
proaches [494], [493], [435], [436], [169] that deal with these issues in describing
polyphonic audio signals.
Paiva, Mendes, and Cardoso [494], [493] proposed a method of obtaining
the melody note sequence by using a model of the human auditory system
[595] as a frequency-analysis front end and applying MIDI-level note track-
ing, segmentation, and elimination techniques. Although the techniques used
differ from the PreFEst method, the basic idea that 'the melody generally
clearly stands out of the background' is the same as the basic PreFEst con-
cept that the FO of the most predominant harmonic structure is considered the
melody. The advantage of this method is that MIDI-level note sequences of the
melody line are generated, while the output of PreFEst is a simple temporal
trajectory of the FO. The method first estimates predominant FO candidates
by using correlograms (see Chapter 8) that represent the periodicities in a
cochleagram (auditory nerve responses of an ear model). It then forms the
temporal trajectories of FO candidates: it quantizes their frequencies to the
closest MIDI note numbers and then tracks them according to their frequency
proximity, where only one-semitone transition is considered continuous. After
this tracking, FO trajectories are segmented into MIDI-level note candidates
by finding a sufficiently long trajectory having the same note number and by
dividing it at clear local minima of its amplitude envelope. Because there still
remain many inappropriate notes, it eliminates notes whose amplitude is too
low, whose duration is too short, or which have harmonically related FOs and
almost same onset and offset times. Finally, the melody note sequence is ob-
tained by selecting the most predominant notes according to heuristic rules.
Since simultaneous notes are not allowed, the method eliminates simultaneous
notes that are less dominant and not in a middle frequency range.
Marolt [435], [436] proposed a method of estimating the melody line by
representing it as a set of short vocal fragments of FO trajectories. This method
is based on the PreFEst method with some modifications: it uses the PreFEst
core to estimate predominant FO candidates, but uses a spectral modelling
synthesis (SMS) front end that performs the sinusoidal modelling and analysis
(see Chapters 1 and 3) instead of the PreFEst front end. The advantage of
this method is that the FO candidates are tracked and grouped into melodic
340 Masataka Goto

fragments (reasonably segmented signal regions that exhibit strong and stable
FO) and these fragments are then clustered into the melody line. The method
first tracks temporal trajectories of the FO candidates (salient peaks) to form
the melodic fragments by using a salient peak tracking approach similar to
the PreFEst back end (though it does not use multiple agents). Because the
fragments belong to not only the melody (lead vocal), but also to different
parts of the accompaniment, they are clustered to find the melody cluster by
using Gaussian mixture models (GMMs) according to their five properties:
Dominance (average weight of a tone model estimated by the EM algo-
rithm) ,
Pitch (centroid of the FOs within the fragment),
Loudness (average loudness of harmonics belonging to the fragment),
Pitch stability (average change of FOs during the fragment), and
Onset steepness (steepness of overall loudness change during the first 50
ms of the fragment).
Eggink and Brown [169] proposed a method of estimating the melody
line with the emphasis on using various knowledge sources, such as knowl-
edge about instrument pitch ranges and interval transitions, to choose the
most likely succession of FOs as the melody line. Unlike other methods, this
method is specialized for a classical sonata or concerto, where a solo melody
instrument can span the whole pitch range, ranging from the low tones of a
cello to a high-pitched flute, so the frequency range limitation used in the
PreFEst method is not feasible. In addition, because the solo instrument does
not always have the most predominant FO, additional knowledge sources are
necessary to extract the melody line. The main advantage of this method is the
leverage provided by knowledge sources, including local knowledge about an
instrument recognition module and temporal knowledge about tone durations
and interval transitions, which are integrated in a probabilistic search. Those
sources can both help to choose the correct FO among multiple concurrent FO
candidates and to determine sections where the solo instrument is actually
present. The knowledge sources consist of two categories, local knowledge and
temporal knowledge. The local knowledge concerning FO candidates obtained
by picking peaks in the spectrum includes
FO strength (the stronger the spectral peak, the higher its likelihood of
being the melody),
Instrument-dependent FO likelihood (the likehhood values of an FO candi-
date in terms of its frequency and the pitch range of each solo instrument,
which are evaluated by counting the frequency of its FO occurrence in
different standard MIDI files), and
Instrument likelihood (the likelihood values of an FO candidate being pro-
duced by each solo instrument, which are evaluated by the instrument
recognition module).
 11 Music Scene Description 341

The instrument recognition module uses trained Gaussian classifiers of the

frequency and power of the first ten harmonic components, their deltas, and
their delta-deltas, which are taken from the spectrum for each FO candidate.
On the other hand, the temporal knowledge concerning tone candidates ob-
tained by connecting FO candidates includes
Instrument-dependent interval likelihood (the likelihood values of an inter-
val transition between succession tones, which are evaluated by counting
the frequency of its interval occurrence in different standard MIDI files),
and
Relative tone usage (measures related to tone durations between successive
tones, which are used to penalize overlapped tones).
These knowledge sources are combined to find the most likely 'path' of the
melody through the space of all FO candidates in time. Since the melody path
occasionally follows the accompaniment, additional postprocessing is done to
eliminate sections where the solo instrument is actually silent.
While the above methods deal with the melody line, Hainsworth and
Macleod [264] proposed a method of obtaining the bass note sequence by
maintaining multiple hypotheses. The method first extracts the onset times
of bass notes by picking peaks of a smoothed temporal envelope of a total
power below 200 Hz. It then generates hypotheses regarding the FO of each
extracted note; the FO and amplitude of each hypothesis are estimated by
fitting a quadratic polynomial to a large amplitude peak and subtracting it
from the spectrum. The first four harmonic components of those hypotheses
are tracked over time by using a comb-filter-like analysis. Finally, the method
selects the most likely hypothesis for each onset on the basis of its duration and
the amplitude of harmonic components and further tidies up these hypotheses
by removing inappropriate overlaps and relatively low amplitude notes.

11.3 Estimating Beat Structure

Beat tracking (including measure or bar line estimation) is defined as the
process of organizing musical audio signals into a hierarchical beat struc-
ture (including beat and measure levels). It is also an important initial step
in the computational modelling of music understanding because the beat is
fundamental, for both trained and untrained listeners, to the perception of
Western music. As described in Section 11.7.2 and Section 4.1, p. 101, there
are many applications such as music-synchronized computer graphics, stage
lighting control, video/audio synchronization, and human-computer improvi-
sation in live ensembles.
Various methods for estimating the beat structure are described in detail in
Chapter 4. Here, the synergy between the estimation of the hierarchical beat
structure, drum patterns, and chord changes is briefly discussed. This syn-
ergy is exploited in a real-time beat-tracking system developed by Goto and
342 Masataka Goto

Muraoka [235], [220], [221]. The estimation of the hierarchical beat structure,
especially the measure (bar line) level, requires the use of musical knowledge
about drum patterns and chord changes; on the other hand, drum patterns and
chord changes are difficult to estimate without referring to the beat structure
of the beat level (quarter note level). The system addresses this issue by lever-
aging the integration of top-down and bottom-up processes (Fig. 11.6) under
the assumption that the time signature of an input song is 4/4. The system
first obtains multiple possible hypotheses of provisional beat times (quarter-
note-level beat structure) on the basis of onset times without using musical
knowledge about drum patterns and chord changes. Because the onset times
of the sounds of bass drum and snare drum can be detected by a bottom-up
frequency analysis described in Section 5.2.3, p. 137, the system makes use
of the provisional beat times as top-down information to form the detected
onset times into drum patterns whose grid is aligned with the beat times. The
system also makes use of the provisional beat times to detect chord changes
in a frequency spectrum without identifying musical notes or chords by name.
The frequency spectrum is sliced into strips at the beat times and the domi-
nant frequencies of each strip are estimated by using a histogram of frequency
components in the strip [240]. Chords are considered to be changed when the
dominant frequencies change between adjacent strips. After the drum patterns
and chord changes are obtained, the higher-level beat structure, such as the
measure level, can be estimated by using musical knowledge regarding them.

11.4 Estimating Drums

The detection of the onset times of drum sounds is important because the
basic rhythms of popular music pieces including drum sounds are mainly
characterized by drum performances. As described in Section 11.7.1, there
are many applications such as rhythm-based music information retrieval and
genre classification.
Various methods for detecting drum sounds are described in detail in
Chapter 5.

11.5 Estimating Chorus Sections and Repeated Sections

Chorus ('hook' or 'refrain') sections of popular music are the most represen-
tative, uplifting, and prominent thematic sections in the music structure of
a song, and human listeners can easily understand where the chorus sections
are because these sections are the most repeated and memorable portions of a
song. Automatic detection of chorus sections is essential for the computational
modelling of music understanding and is useful in various practical applica-
tions. In music browsers or music retrieval systems, it enables a listener to
quickly preview a chorus section as a 'music thumbnail' (a musical equivalent
 11 Music Scene Description 343

Beat Structure^^^^

Musical Decision
(higher level, reliability)

Beat Prediction
(quarter-note level)

^nset-time Vector

Onset-time
Vectonzer

Fig. 11.6. Synergy between the estimation of the hierarchical beat structure, drum
patterns, and chord changes. Drum patterns and chord changes are obtained, at
'Higher Analysis' in the figure, by using provisional beat times as top-down infor-
mation. The hierarchical beat structure is then estimated, at 'Musical Decision'
in the figure, by using the drum patterns and chord changes. A drum pattern is
represented by the temporal pattern of a bass drum (BD) and a snare drum (SD).

of an image thumbnail) to find a desired song. It can also provide novel music
listening interfaces for end users as described in Section 11.7.3.
To detect chorus sections, typical approaches do not rely on prior informa-
tion regarding acoustic features unique to choruses but focus on the fact that
chorus sections are usually the most repeated sections of a song. They thus
adopt the following basic strategy: detect similar sections that repeat within a
musical piece (such as a repeating phrase) and output those that appear most
often. On entering the 2000s, this strategy has led to methods for extracting a
single segment from several chorus sections by detecting a repeated section of
a designated length as the most representative part of a musical piece [417],
[27], [103]; methods for segmenting music, discovering repeated structures,
or summarizing a musical piece through bottom-up analyses without assum-
ing the output segment length [110], [111], [512], [516], [23], [195], [104], [82],
[664], [420]; and a method for exhaustively detecting all chorus sections by
determining the start and end points of every chorus section [224].
Although this basic strategy of finding sections that repeat most often is
simple and effective, it is diflftcult for a computer to judge repetition because it
is rare for repeated sections to be exactly the same. The following summarizes
the main problems that must be addressed in finding music repetition and
determining chorus sections.
344 Masataka Goto

Problem 1: Extracting acoustic features and calculating their similarity

Whether a section is a repetition of another must be judged on the basis
of the similarity between the acoustic features obtained from each frame
or section. In this process, the similarity must be high between acoustic
features even if the accompaniment or melody line changes somewhat in
the repeated section (e.g., the absence of accompaniment on bass and/or
drums after repetition). That is, it is necessary to use features that capture
useful and invariant properties.
Problem 2: Finding repeated sections
A pair of repeated sections can be found by detecting contiguous tem-
poral regions having high similarity. However, the criterion establishing
how high similarity must be to indicate repetition depends on the song.
For a song in which repeated accompaniment phrases appear very often,
for example, only a section with very high similarity should be considered
the chorus section repetition. For a song containing a chorus section with
accompaniments changed after repetition, on the other hand, a section
with somewhat lower similarity can be considered the chorus section rep-
etition. This criterion can be easily set for a small number of specific songs
by manual means. For a large open song set, however, the criterion should
be automatically modified based on the song being processed.
Problem 3: Grouping repeated sections
Even if many pairs of repeated sections with various lengths are obtained,
it is not obvious how many times and where a section is repeated. It
is therefore necessary to organize repeated sections that have common
sections into a group. Both ends (the start and end points) of repeated
sections must also be estimated by examining the mutual relationships
among various repeated sections. For example, given a song having the
structure (A B C B C C), the long repetition corresponding to (B C) would
be obtained by a simple repetition search. Both ends of the C section in (B
C) could be inferred, however, from the information obtained regarding
the final repetition of C in this structure.
Problem 4' Detecting modulated repetition
Because the acoustic features of a section generally undergo a significant
change after modulation (key change; see Section 1.1, p. 7), similarity
with the section before modulation is low, making it difficult to judge
repetition. The detection of modulated repetition is important since mod-
ulation sometimes occurs in chorus repetitions, especially in the latter half
of a song.^
Problem 5: Selecting chorus sections
Because various levels of repetition can be found in a musical piece, it is
necessary to select a group of repeated sections corresponding to chorus

^Masataka Goto's survey of Japan's popular music hit chart (top 20 singles
ranked weekly from 2000 to 2003) showed that modulation occurred in chorus rep-
etitions in 152 songs (10.3%) out of 1481.
 11 Music Scene Description 345

sections. A simple selection of the most repeated sections is not always

appropriate though. For example, another section such as verse A is oc-
casionally repeated more often than chorus sections.
Regarding the above repetition-based methods, the following sections
mainly describe a method called RefraiD (Refrain Detection method) [224]
and briefly introduce techniques used in the other methods in each relevant
section and Section 11.5.6. Since the RefraiD method addresses all of the above
problems and detects all chorus sections in a popular music song regardless
of whether a key change occurs, it is suitable for music scene description. Fig-
ure 11.7 shows the process flow of the RefraiD method. First, a 12-dimensional
feature vector called a chroma vector^ which is robust with respect to changes
of accompaniments, is extracted from each frame of an input audio signal and
then the similarity between these vectors is calculated (solution to Problem
1). Each element of the chroma vector corresponds to one of the 12 pitch
classes (C, C # , D, ..., B) and is the sum of the magnitude spectrum at fre-
quencies of its pitch class over six octaves. Pairs of repeated sections are then
listed (found) using an adaptive repetition-judgement criterion which is con-
figured by an automatic threshold selection method based on a discriminant
criterion (solution to Problem 2). To organize common repeated sections into
groups and to identify both ends of each section, the pairs of repeated sec-
tions are integrated (grouped) by analysing their relationships over the whole
song (solution to Problem 3). Because each element of a chroma vector corre-
sponds to a different pitch class, a before-modulation chroma vector is close
to the after-modulation chorus vector whose elements are shifted (exchanged)
by the pitch difference of the key change. By considering 12 kinds of shift
(pitch differences), 12 sets of the similarity between non-shifted and shifted
chroma vectors are then calculated, pairs of repeated sections from those sets
are listed, and all of them are integrated (solution to Problem 4)- Finally,
the chorus measure^ which is the possibility of being chorus sections for each
group, is evaluated (solution to Problem 5)^ and the group of chorus sections
with the highest chorus measure as well as other groups of repeated sections
are output (Fig. 11.8).

11.5.1 Extracting Acoustic Features and Calculating Their

Similarity

The following acoustic features, which capture pitch and timbral features of
audio signals in different ways, were used in various methods: chroma vectors
[224], [27], [110], [111], mel-frequency cepstral coefficients (MFCC) [417], [103],
[23], [195], [104], (dimension-reduced) spectral coefficients [103], [195], [104],
[82], [664], pitch representations using FO estimation or constant-Q filterbanks
[110], [111], [82], [420], and dynamic features obtained by supervised learning
[512], [516].
346 Masataka Goto

Audio
^ signals
^ O
o
[ Extracting 12-dimensional chroma vectorsj D
CO
Q!
0

[ Calculating similarity between chroma vectors ]

Listing repeated sections

Integrating repeated sections ]

CD
[integrating repeated sections with modulation]

Selecting chorus sections

^ _ 4
I List of chorus sections | | List of repeated sections |

Fig. 11.7. Overview of RefraiD (Refrain Detection method) for detecting all chorus
sections with their start and end points while considering modulations (key changes).

Fig. 11.8. An example of chorus sections and repeated sections detected by the
RefraiD method. The horizontal axis is the time axis (in seconds) covering the
entire song. The upper window shows the power. On each row in the lower window,
coloured sections indicate similar (repeated) sections. The top row shows the list
of the detected chorus sections, which were correct for this song (RWC-MDB-P-
2001 No. 18 of the RWC Music Database [229], [227]) and the last of which was
modulated. The bottom five rows show the list of various repeated sections (only
the five longest repeated sections are shown). For example, the second row from the
top indicates the structural repetition of 'verse A => verse B => chorus'; the bottom
row with two short coloured sections indicates the similarity between the 'intro' and
'ending'.

Pitch Feature: C h r o m a Vector

T h e chroma vector is a perceptually motivated feature vector using the con-

cept of chroma in Shepard's helix representation of musical pitch perception
[584]. According to Shepard [584], the perception of pitch with respect to a
musical context can be graphically represented by using a continually cyclic
helix t h a t has two dimensions, chroma and height^ as shown at t h e right of
Fig. 11.9. C h r o m a refers to the position of a musical pitch within an octave
 11 Music Scene Description 347

that corresponds to a cycle of the hehx; i.e., it refers to the position on the
circumference of the hehx seen from directly above. On the other hand, height
refers to the vertical position of the helix seen from the side (the position of
an octave).
Figure 11.9 shows an overview of calculating the chroma vector used in
the RefraiD method [224]. This represents magnitude distribution on the
chroma that is discretized into twelve pitch classes within an octave. The 12-
dimensional chroma vector v{t) is extracted from the magnitude spectrum,
^p{y^t) at the log-scale frequency v at time t, calculated by using the short-
time Fourier transform (STFT). Each element of v(t) corresponds to a pitch
class c (c = 1, 2 , . . . , 12) in the equal temperament and is represented as Vc{t)'.
QctH poo
^c{t)= Y, / ^PhA^)%{^.t)dv, (11.29)
/i = OctL * ^ ~ ^

The BPFc,/i(z/) is a bandpass filter that passes the signal at the log-scale centre
frequency Fc^h (in cents) of pitch class c (chroma) in octave position h (height),
where
Fc^h = 1200/1 -f 100(c - 1). (11.30)
The BPFc,/i(z/) is defined using a Manning window as follows:

BPFe,.(.) = I ( l - COS ? ! f c i | ^ L Z i M ) ) , , e [0,200]. (11.31)

This filter is applied to octaves from OctL to Octn- In Goto's implementation

[224], an STFT with a 256 ms Hanning window^ shifted by 80 ms is calculated
for audio signals sampled at 16 kHz, and the OctL and Octn are respectively
3 and 8, covering six octaves (130 Hz to 8 kHz).
There are variations in how the chroma vector is calculated. For example,
Bartsch and Wakefield [27] developed a technique where each STFT bin of
the log-magnitude spectrum is mapped directly to the most appropriate pitch
class, and Dannenberg and Hu [110], [111] also used this technique. A similar
continuous concept was called the chroma spectrum [655].
There are several advantages to using the chroma vector. Because it cap-
tures the overall harmony (pitch-class distribution), it can be similar even if
accompaniments or melody lines are changed to some degree after repetition.
In fact, the chroma vector is effective for identifying chord names [201], [678],
[679], [583], [684]. The chroma vector also enables modulated repetition to be
detected as described in Section 11.5.4.

Timbral Feature: M F C C and Dynamic Features

While the chroma vectors capture pitch-related content, the MFCCs (see
Section 2.1.3, p. 25 for a presentation of MFCCs) typically used in speech
^The window length is determined to obtain good frequency resolution in a low-
frequency region.
348 Masataka Goto

STFT magnitude spectrum

log-scale frequency
I ^
LULu.
130 Hz
.kA .n i l l *< L y k hrf
8 kHz

Summation
Chroma vector over six octaves
B^v^chroma^_
C# D
Helix representation
of pitch perception
ICIC#IDID#IEIFIF#IGIG#I AIA#IBI
Fig. 11.9. Overview of calculating a 12-dimensional chroma vector. The magnitude
at six different octaves is summed into just one octave which is divided into 12
log-spaced divisions corresponding to pitch classes. Shepard's helix representation
of musical pitch perception [584] is shown at the right.

recognition capture spectral content and general pitch range, and are useful
for finding timbral or 'texture' repetitions. Dynamic features [512], [516] are
more adaptive spectral features that are designed for music structure discov-
ery through a supervised learning method. Those features are selected from
the spectral coefiicients of a filterbank output by maximizing the mutual in-
formation between the selected features and hand-labelled music structures.
The dynamic features are beneficial in that they reduce the size of the results
when calculating similarity (i.e., the size of the similarity matrix described in
Section 11.5.1) because the frame shift can be longer (e.g., 1 s) than for other
features.

Calculating Similarity

Given a feature vector such as the chroma vector or MFCC at every frame,
the next step is to calculate the similarity between feature vectors. Various
distance or similarity measures, such as the Euclidean distance and the cosine
angle (inner product), can be used for this. Before calculating the similarity,
feature vectors are usually normalized, for example, to a mean of zero and a
standard deviation of one or to a maximum element of one.
In the RefraiD method [224], the similarity r(t, /) between the feature vec-
tors (chroma vectors) v(^) and v{t I) is defined as

1 v{t) v{t-l)
r(t,0 = l - (11.32)
maxc Vc{t) maxc Vc{t I)

where / is the lag and Vc{t) is an element of v(^) (11.29). Since the denominator
A/12 is the length of the diagonal line of a 12-dimensional hypercube with edge
length 1, r{tj) satisfies 0 < r{tj) < 1.
 11 Music Scene Description 349

r(t,l)

t(time)
lAlAlBlBlAlA

(a) Similarity matrix (b) Time-lag triangle

Fig. 11.10. An idealized example of a similarity matrix and time-lag triangle drawn
from the same feature vectors of a musical piece consisting of four 'A' sections and
two 'B' sections. The diagonal line segments in the similarity matrix or horizontal
line segments in the time-lag triangle, which represent similar sections, appear when
short-time pitch features like chroma vectors are used.

For the given 36-dimensional feature vectors of a constant-Q filterbank

o u t p u t with centre frequencies at 36 tempered semitones in 3 octaves, Lu,
Wang, and Zhang [420] introduced an original distance measure t h a t empha-
sizes melody similarity and suppresses timbre similarity. This measure does
not depend on the norm of the difference between the 36-dimensional fea-
ture vectors, but on the structure of it. It considers how the peak intervals
in the difference conform to harmonic relationships such as perfect fifth and
octave.

11.5.2 Finding R e p e a t e d Sections

By using the same similarity measure r ( t , / ) , two equivalent representations

can be obtained: a similarity matrix [103], [110], [ H I ] , [195], [104], [664] and
a time-lag triangle (or time-lag matrix) [224], [27], [516], [420], as shown in
Fig. 11.10. For the similarity matrix, the similarity s{t,u) between feature
vectors v{t) and v(w),
s{t,u)^ r{t,t-u), (11.33)

is drawn within a square in the two-dimensional {t-u) space.^ For the time-
lag triangle, the similarity r(t,/) between feature vectors v ( t ) and v ( t /) is
drawn within a right-angled isosceles triangle in the two-dimensional time-lag
[t-l) space. If a nearly constant t e m p o can be assumed, each pair of simi-
lar sections is represented by two non-central diagonal line segments in the

As described in Section 4.6, p. 112, the similarity matrix can also be used to
examine rhythmic structure.
350 Masataka Goto

similarity matrix or a horizontal line segment in the time-lag triangle. Because

the actual r{tj) obtained from a musical piece is noisy and ambiguous, it is
not a straightforward task to detect these line segments.
T h e RefraiD method [224] finds all horizontal fine segments (contiguous
regions with high r(t,/)) in the time-lag triangle by evaluating Raii(t,/), the
possibility of containing line segments at the lag / at the current time t (e.g.,
at the end of a song^) as follows (Fig. 11.11):^

1 *

T= l

Before this calculation, r(f, /) is normalized by subtracting a local mean value

while removing noise and emphasizing horizontal lines. In more detail, given
each point r(t, /) in the time-lag triangle, six-directional local mean values
along the right, left, upper, lower, upper right, and lower left directions start-
ing from the point r(f, /) are calculated, and the maximum and minimum are
obtained. If the local mean along the right or left direction takes the maximum,
r(t, /) is considered part of a horizontal line and emphasized by subtracting
the minimum from r(t, / ) . Otherwise, r(t, /) is considered noise and suppressed
by subtracting the maximum from r(t, /); noise tends to appear as fines along
the upper, lower, upper right, and lower left directions.
T h e method then picks u p each peak in Raii(^,0 along the lag / after
smoothing Raii(^, 0 with a moving average filter along the lag and removing
a global drift (bias) caused by cumulative noise in r(t,/)^ from Raii(t,/). The
m e t h o d next selects only high peaks above a threshold to search the line seg-
ments. Because this threshold is closely related to the repetition-judgement
criterion which should be adjusted for each song, an automatic threshold selec-
tion method based on a discriminant criterion [491] is used. W h e n dichotomiz-
ing the peak heights into two classes by a threshold, the optimal threshold
is obtained by maximizing the discriminant criterion measure defined by the
following between-class variance:

a% = uJiuj2{fii - fi2f, (11.35)

where ui and LJ2 are the probabilities of class occurrence (number of peaks in
each class/total number of peaks), and /ii and /12 are the means of the peak
heights in each class.

^ Rail (^5 0 is evaluated along with the real-time audio input for a real-time system
based on RefraiD. On the other hand, it is evaluated at the end of a song for a non-
real-time off-line analysis.
^This can be considered the Hough transform where only horizontal lines are
detected: the parameter (voting) space Rail(^5 0 i^ therefore simply one dimensional
along /.
^Because the similarity r(T,/) is noisy, its sum Raii(^, 0 tends to be biased:
the longer the summation period for Rail(^5 05 ^he higher the summation result by
(11.34).
 11 Music Scene Description 351

Line segments cxHTesponding

t) repetition of half of C
Line segments conBsponding A
t) repetition of C

Line segment corresponding \

to repetition of ABCC J N-

Line segment \ i _I
ccn-esponding to i
repetition of C ^-^

A: verse A
B: verse B r(t,l)
C:chorus

0 BiClCl |A|B|C|C|CT^^^^^'^^>^^-^^^^'^^

Fig. 11.11. A sketch of line segments, the similarity r(t,/) in the time-lag triangle,
and the possibility Raii(^, 0 ^^ containing line segments at lag /.

similarity 1
c o ^m region above a threshold |
Is

iwjj^ 'T*%#fl
Q E
^^\
is -c
_co o I
E
CO

timex

similarity 1
c o region above a threshold |

|i
CD CO

wH m%f^
-Q E
_co o
E
^
timex

Fig. 11.12. Examples of the similarity r[r^ h) at high-peak lags h. The bottom hor-
izontal bars indicate the regions above an automatically adjusted threshold, which
means they correspond to line segments.

For each picked-up high peak with lag / i , the line segments are finally
searched on the one-dimensional function r(r, h) [h <T <t). After smoothing
r(T,/i) using a moving average filter, the m e t h o d obtains line segments on
352 Masataka Goto

which the smoothed r(r, Zi) is above a threshold (Fig. 11.12). This threshold
is also adjusted through the automatic threshold selection method.
Instead of using the similarity matrix and time-lag triangle, there are other
approaches that do not explicitly find repeated sections. To segment music,
represent music as a succession of states (labels), and obtain a music thumbnail
or summary, these approaches segment and label (i.e., categorize) contiguous
frames (feature vectors) by using clustering techniques [417] or ergodic hidden
Markov models (HMMs) [417], [512], [516] (HMMs are introduced on p. 63 of
this volume).

11.5.3 Grouping Repeated Sections

Since each line segment in the time-lag triangle indicates just a pair of re-
peated sections, it is necessary to organize into a group the line segments that
have common sectionsi.e., overlap in time. When a section is repeated N
times {N > 3), the number of line segments to be grouped together should
theoretically be N{N l ) / 2 if all of them are found in the time-lag triangle.
Aiming to exhaustively detect all the repeated (chorus) sections appearing
in a song, the RefraiD method groups line segments having almost the same
section while redetecting some missing (hidden) line segments not found in the
bottom-up detection process (described in Section 11.5.2) through top-down
processing using information on other detected fine segments. In Fig. 11.11,
for example, two line segments corresponding to the repetition of the first
and third C and the repetition of the second and fourth C, which overlap
with the long line segment corresponding to the repetition of ABCC, can be
found even if they were hard to find in the bottom-up process. The method
also appropriately adjusts the start and end times of line segments in each
group because they are sometimes inconsistent in the bottom-up line segment
detection.

11.5.4 Detecting Modulated Repetition

The processes described above do not deal with modulation (key change), but
they can easily be extended to it. A modulation can be represented by the
pitch difference of its key change, C ( 0 - l r - - - . l l ) , which denotes the number
of tempered semitones. For example, C = 9 means the modulation of nine
semitones upward or the modulation of three semitones downward. One of the
advantages of the 12-dimensional chroma vector v{t) is that a transposition
amount (" of the modulation can naturally correspond to the amount by which
its 12 elements are shifted (rotated). When v{t) is the chroma vector of a
certain performance and v(^)' is the chroma vector of the performance that is
modulated by C semitones upward from the original performance, they tend
to satisfy
v(^) ^ S^v(t)"^, (11.36)
 11 Music Scene Description 353

where S is a 12-by-12 shift matrix defined by

/O 1 0 0\
0 0 1 0 0
S = (11.37)
0... 0 1 0
0... 0 1
\1 0 0/
To detect modulated repetition by using this feature of chroma vectors
and considering 12 destination keys, the RefraiD method [224] calculates 12
kinds of extended similarity as follows:

1 Sv(i) v(i-0
r^itj) (11.38)
maxc Vc{t) maxc Vc{t I)

Starting from each r(^{t, /), the processes of finding and grouping the repeated
sections are performed again. Non-modulated and modulated repeated sec-
tions are then grouped if they share the same section.

11.5.5 Selecting Chorus Sections

A group corresponding to the chorus sections is finally selected from groups of

repeated sections (line segments). In general, a group that has many and long
repeated sections tends to be the chorus sections. In addition to this property,
the RefraiD method evaluates the chorus measure^ which is the possibility
of being chorus sections for each group, by considering the following three
heuristic rules with a focus on popular music:
1. The length of the chorus has an appropriate, allowed range (7.7 to 40 s in
Goto's implementation).
2. When there is a repeated section that is long enough to likely correspond
to the repetition of a long section like (verse A ^ verse B =^ chorus) x
2, the chorus section is likely to be at the end of that repeated section.
3. Because a chorus section tends to have two half-length repeated subsec-
tions within its section, a section having those subsections is likely to be
the chorus section.
The group that maximizes the chorus measure is finally selected as the chorus
sections.

11.5.6 Other Methods

Since the above sections mainly describe the RefraiD method [224] with the
focus on detecting all chorus sections, this section briefly introduces other
methods [417], [27], [103], [110], [111], [512], [516], [23], [195], [104], [82], [664],
354 Masataka Goto

[420] that aim at music thumbnailing, music segmentation, structure discovery,

or music summarization.
Several methods for detecting the most representative part of a song for
use as a music thumbnail have been studied. Logan and Chu [417] developed
a method using clustering techniques and hidden Markov models (HMMs) to
categorize short segments (1 s) in terms of their acoustic features, where the
most frequent category is then regarded as a chorus. Bartsch and Wakefield
[27] developed a method that calculates the similarity between acoustic fea-
tures of beat-length segments obtained by beat tracking and finds the given-
length segment with the highest similarity averaged over its segment. Cooper
and Foote [103] developed a method that calculates a similarity matrix of
acoustic features of short frames (100 ms) and finds the given-length segment
with the highest similarity between it and the whole song. Note that these
methods assume that the output segment length is given and do not identify
both ends of a repeated section.
Music segmentation or structure discovery methods where the output seg-
ment length is not assumed have also been studied. Dannenberg and Hu [110],
[111] developed a structure discovery method of clustering pairs of similar seg-
ments obtained by several techniques such as efficient dynamic programming
or iterative greedy algorithms. This method finds, groups, and removes sim-
ilar pairs from the beginning to group all the pairs. Peeters, La Burthe, and
Rodet [512], [516] developed a supervised learning method of modelling dy-
namic features and studied two structure discovery approaches: the sequence
approach of obtaining repetitions of patterns and the state approach of ob-
taining a succession of states. The dynamic features are selected from the
spectrum of a filterbank output by maximizing the mutual information be-
tween the selected features and hand-labelled music structures. Aucouturier
and Sandler [23] developed two methods for finding repeated patterns in a
succession of states (texture labels) obtained by HMMs. They used two image
processing techniques, the kernel convolution and Hough transform, to detect
line segments in the similarity matrix between the states. Foote and Cooper
[195], [104] developed a method of segmenting music by correlating a kernel
along the diagonal of the similarity matrix, and clustering the obtained seg-
ments on the basis of the self-similarity of their statistics. Chai and Vercoe
[82] developed a method of detecting segment repetitions by using dynamic
programming, clustering the obtained segments, and labelling the segments
based on heuristic rules such as the rule of first labelling the most frequent
segments, removing them, and repeating the labelling process. Wellhausen
and Crysandt [664] studied the similarity matrix of spectral envelope features
defined in the MPEG-7 descriptors and a technique of detecting non-central
diagonal line segments. Lu, Wang, and Zhang [420] developed a method of
analysing all repeated sections by using a structure-based distance measure
that emphasizes pitch similarity over timbral similarity. Their method also es-
timates the tempo of a song and discriminates between vocal and instrumental
sections to facilitate music structure analysis.
 11 Music Scene Description 355

11.6 Evaluation Issues

To evaluate automatic music scene description methods, it is necessary to

label musical pieces in an adequate-size music database with their correct
descriptions (metadata). This labelling task is time consuming and trouble-
some. More seriously, there was no available common music database with
correct metadata since most musical pieces used by researchers are generally
copyrighted and cannot be shared by other researchers.
But since 2000, a copyright-cleared music database, called the RWC (Real
World Computing) Music Database [229], [230], [227], was developed and has
been available to researchers as a common foundation for research. It contains
six original collections: the Popular Music Database (100 pieces), Royalty-
Free Music Database (15 pieces). Classical Music Database (50 pieces). Jazz
Music Database (50 pieces). Music Genre Database (100 pieces), and Musical
Instrument Sound Database (50 instruments). For all 315 musical pieces, audio
signals, standard MIDI files, and text files of lyrics were prepared. For the
50 instruments, individual sounds at half-tone intervals were captured. This
database has been distributed to researchers around the world and has already
been widely used. For musical instrument sounds, there are other databases
released for public use: the McGill University Master Samples [487] and the
University of Iowa Musical Instrument Samples [198]. Musical pieces licensed
under a Creative Commons license can also be used for evaluation purposes.
To establish benchmarks (evaluation frameworks) for music scene descrip-
tion by labelling copyright-cleared musical pieces with correct descriptions, a
multipurpose music-scene labelling editor (metadata editor) was also devel-
oped [225]. It enables a user to hand-label a musical piece with music scene
descriptions shown in Fig. 11.1. The editor can deal with both audio files and
standard MIDI files and supports interactive audio/MIDI playback while edit-
ing. Along a wave or MIDI piano-roll display it shows subwindows in which
any selected descriptions can be displayed and edited. To facilitate the sup-
port of various descriptions, its architecture is based on a plug-in system in
which an external module for editing each description is installed as plug-
in software. As a first step, the RefraiD method was evaluated by using the
chorus section metadata for 100 songs of the RWC Music Database: Popular
Music (80 of the 100 songs were correctly detected) [224].

11.7 Applications of Music Scene Description

Music scene description methods that can deal with real-world audio signals of
musical pieces sampled from CD recordings have various practical applications
such as music information retrieval, music-synchronized computer graphics,
and music listening stations. The following sections introduce these applica-
tions.
356 Masataka Goto

Fig. 11.13. Virtual dancer 'Cindy'.

11.7.1 Music Information Retrieval

Music scene description contributes to content-based music information re-

trieval since it can provide various acoustical metadata (annotations) of mu-
sical pieces. For example, the automatic melody estimation described in Sec-
tion 11.2 is useful for query by humming (QBH) [323], [207], [604], [484], [603],
[507], [605], [585], [301], [109] which enables a user to retrieve a musical piece
by humming or singing its melody: a QBH database consisting of audio sig-
nals of musical pieces can be indexed using their melody lines. Moreover, the
description of chorus sections (Section 11.5) can increase the efficiency and
precision of QBH by enabling a QBH system to match a query with only the
chorus sections.
Temporal or rhythmic descriptions such as beat structure, tempo, and
drums (Sections 11.3 and 11.4) are also useful for retrieving musical pieces on
the basis of rhythm and tempo. Indexing musical pieces using drum descrip-
tions, for example, will enable a user to retrieve music by voice percussion or
beat boxing (verbalized expression of drum sounds by voice) [479], [326].
In addition, various music scene descriptions facilitate the computation of
similarity between musical pieces. Similarity measures based on music scene
descriptions enable a user to use musical pieces themselves as the search key
to retrieve a musical piece having a similar feeling. These measures can also
be used to automatically classify musical pieces into genres or music styles.

11.7.2 Music-Synchronized Computer Graphics

Because the beat tracking described in Section 11.3 and Chapter 4 can be
used to automate the time-consuming tasks that must be done to synchronize
events with music, there are various applications. In fact, Goto and Muraoka
[235], [220], [221] developed a real-time system that displays virtual dancers
and several graphic objects whose motions and positions change in time to
beats (Fig. 11.13). This system has several dance sequences, each for a different
mood of dance motions. While a user selects a dance sequence manually, the
timing of each motion in the selected sequence is determined automatically
 11 Music Scene Description 357

on the basis of the beat-tracking results. Such a computer graphics system is

suitable for live stage, TV program, and karaoke uses.
Beat tracking also facilitates the automatic synchronization of computer-
controlled stage lighting with the beats in a musical performance. Vari-
ous properties of lightingsuch as colour, brightness, and directioncan be
changed in time to music. In the above virtual dancer system, this was simu-
lated on a computer graphics display with virtual dancers.

11.7.3 Music Listening Station

The automatic chorus section detection described in Section 11.5 enables new
music-playback interfaces that facilitate content-based manual browsing of
entire songs. As an application of the RefraiD method. Goto [226] developed
a music Hstening station for trial listening, called SmartMusicKIOSK. Cus-
tomers in music stores often search out the chorus or 'hook' of a song by
repeatedly pressing the fast-forward button, rather than passively listening to
the music. This activity is not well supported by current technology. SmartMu-
sicKIOSK provides the following two functions to facilitate an active listening
experience by eliminating the hassle of manually searching for the chorus and
making it easier for a listener to find desired parts of a song:
1. ^Jump to chorus' function: automatic jumping to the beginning of sections
relevant to a song's structure
Functions are provided enabling automatic jumping to sections that will
be of interest to listeners. These functions are 'jump to chorus (NEXT
CHORUS button)', 'jump to previous section in song (PREV SECTION
button)', and 'jump to next section in song (NEXT SECTION button)',
and they can be invoked by pushing the buttons shown above in paren-
theses (in the lower window of Fig. 11.14). With these functions, a listener
can directly jump to and listen to chorus sections, or jump to the previous
or next repeated section of the song.
2. 'Music map' function: visualization of song contents
A function is provided to enable the contents of a song to be visuahzed
to help the listener decide where to jump next. Specifically, this function
provides a visual representation of the song's structure consisting of chorus
sections and repeated sections, as shown in the upper window of Fig. 11.14.
While examining this display, the listener can use the automatic jump
buttons, the usual fast-forward/rewind buttons, or a playback slider to
move to any point of interest in the song.
This interface, which enables a listener to look for a section of interest
by interactively changing the playback position, is useful not only for trial
listening but also for more general purposes in selecting and using music.
While entire songs of no interest to a listener can be skipped on conventional
music-playback interfaces, SmartMusicKIOSK is the first interface that allows
the listener to easily skip sections of no interest even within a song.
358 Masataka Goto

F i g . 11.14. SmartMusicKIOSK screen display. The lower window provides content-

based controls allowing a listener to skim rapidly through music as well as common
playback controls. The upper window provides a graphical overview of the music
structure (results of automatic chorus section detection using RWC-MDB-P-2001
No. 18 of the RWC Music Database [229], [227]). The horizontal axis of the upper
window is the time axis covering the entire song; the top row shows chorus sections,
the five lower rows show repeated sections, and the bottom horizontal bar is a
playback slider.

11.8 Conclusion

This chapter has described the music scene description research approach to-
wards developing a system t h a t understands real-world musical audio signals
without deriving musical scores or separating signals. This approach is im-
p o r t a n t from an academic viewpoint because it explores what is essential for
understanding audio signals in a human-like fashion. T h e ideas and techniques
are expected to be extended to not only music signals b u t also general audio
signals including music, speech, environmental sounds, and mixtures of them.
Traditional speech recognition frameworks have been developed for dealing
with only monophonic speech signals or a single-pitch sound with background
noise, which should be removed or suppressed without considering their rela-
tionship. Research on understanding musical audio signals is a good starting
point for creating a new framework for understanding general audio signals,
because music is polyphonic, temporally structured, and complex, yet still
well organized. In particular, relationships between various simultaneous or
successive sounds are important and unique to music. This chapter, as well as
other chapters in this book, will contribute to such a general framework.
T h e music scene description approach is also important from industrial or
application viewpoints since end users can now easily 'rip' audio signals from
CDs, compress and store t h e m on a personal computer, load a huge number of
 11 Music Scene Description 359

songs onto a portable music player, and listen to them anywhere and anytime.
These users want to retrieve and listen to their favourite music or a portion
of a musical piece in a convenient and flexible way. Reflecting these demands,
the target of processing has expanded from the internal content of individual
musical pieces to entire musical pieces and even sets of musical pieces [233].
While the primary target of music scene description is the internal content of
a piece, the obtained descriptions are useful for dealing with sets of musical
pieces as described in Section 11.7.1. The more accurate and detailed we can
make the obtained music scene descriptions, the more advanced and intelligent
music applications and interfaces will become.
Although various methods for detecting melody and bass lines, tracking
beats, detecting drums, and finding chorus sections have been developed and
successful results have been achieved to some extent, there is much room for
improving these methods and developing new ones. For example, in general
each method has been researched independently and implemented separately.
An integrated method exploiting the relationships between these descriptions
will be a promising next step. Other music scene descriptions apart from
those described in this chapter should also be investigated in the future. Ten
years ago it was considered too difficult for a computer to obtain most of the
music scene descriptions described here, but today we can obtain them with a
certain accuracy. I look forward to experiencing further advances in the next
ten years.
12
Singing Transcription

Matti Ryynanen

Institute of Signal Processing, Tampere University of Technology

Korkeakoulunkatu 1, 33720 Tampere, Finland
Matti. RyynanenQtut. f i

12.1 Introduction
Singing refers to the act of producing musical sounds with the human voice,
and singing transcription refers to the automatic conversion of a recorded
singing signal into a parametric representation (e.g., a MIDI file) by apply-
ing signal-processing methods. Singing transcription is an important topic
in computational music-content analysis since it is the most natural way of
human-computer interaction in the musical sense: even a musically untrained
subject is usually able to hum the melody of a piece. This chapter introduces
the singing transcription problem and presents an overview of the main ap-
proaches to solve it, including the current state-of-the-art singing transcription
systems.
During the last ten years, the rapid growth of digital music databases has
challenged researchers to develop natural user interfaces for accessing them
by using the singing voice. Consequently, most of the research on singing
transcription has been conducted in the context of query-by-humming systems
where singing transcription acts as a front end. After converting a singing
signal into a notated query, music pieces corresponding to the query can be
retrieved from the database. However, singing transcription enables a wide
range of other applications as well, including singing-input functionalities in
applications such as computer games or singing tutors, automatic tools for
annotating large corpora of singing, audio editor applications for professional
music production, and naturally, applications that convert singing signals into
musical scores.

12.1.1 Problem Formulation and the Scope of This Chapter

The singing transcription problem is here formulated as follows.
Given the acoustic waveform of a single-voice singing performance,
produce a sequence of notes and rests which is melodically and rhyth-
mically as close to the performance as possible.
362 Matti Ryynanen

The sequence of notes and rests forms a parametric representation that is

considered as the transcription of the singing signal. A note has an identifi-
able pitch, a beginning (onset) and an ending (offset) time, and a duration
determined by these. Rests are silent moments in music. The fundamental
frequency (FO) of a note is measured in Hertz (Hz) units and corresponds to
the pitch of a note (see Chapter 1 for the definitions of pitch and FO). The
ratio of the fundamental frequencies of two notes is referred to as an interval.
In particular, an interval with the ratio 1 : 2 is called an octave^ which is di-
vided into twelve notes in Western music. This leads to the FO ratio 1 : 2^/^^
between adjacent notes, which is called a semitone.
A note pitch label identifies the standard FO of a note either with a note
name (e.g., A4) or with an integer MIDI note number. The latter is defined
for a note with a fundamental frequency FQ by

MIDI note number = 69 -^ 12 logo ( T T ; ^ ) , (12.1)

^ V440Hzy ^ ^
where 69 and 440 Hz correspond to the MIDI note number and the funda-
mental frequency of the note A4, respectively. The term absolute tuning refers
to a standard tuning where the note A4 has a FO value of 440 Hz, and the
FOs of all the other notes are related to it. Equation (12.1) provides a musi-
cally convenient way of representing an arbitrary FO value in semitone units
by omitting the rounding to an integer. The term note labelling refers to the
assignment of a pitch label for a note (e.g., an integer MIDI note number).
The term note segmentation refers to a process where the onset and offset
time of a note are determined.
The biggest difficulty in singing transcription lies in the conversion of a
continuous FO curve into note pitch labels. Simply rounding FO estimates
into MIDI note numbers produces poor results since the FO curve may ex-
hibit large deviations from the nominal note pitches. Singing performances
are typically far from perfect and usually there is no straightforward one-to-
one correspondence between a FO curve and a transcribed note. In addition,
it cannot be assumed that singing is performed in absolute tuning. In this
respect, singing transcription is particularly challenging as compared to the
transcription of other musical instruments which usually produce more sta-
ble FO curves for each note. The relationship between singing sounds and
note pitches is discussed in more detail in Section 12.2.
This chapter concentrates on the quantization of note pitches rather than
their onset times and durations, although the common musical notation ex-
presses also durations with discrete labels. In this sense, the aim here is to
produce a piano-roll or MIDI-type representation of singing rather than the
actual score of the performance. There exist algorithms for the time quan-
tization of notes for producing score-type representations, such as work by
Cemgil et al. [75], [77], which can be used to post-process the transcription
result. See Chapter 4 for a discussion on time quantization.
 12 Singing Transcription 363

SINGING ^
FEATURE
PRE-PROCESSING
SIGNAL EXTRACTION

ENHANCED
i ACOUSTIC AND
-^ NOTE SEGMENTATION -*
NOTE - POST-PROCESSING MUSICOLOGICAL
AND LABELING
SEQUENCE MODELS
NOTE SEQUENCE

Fig. 12.1. A flow chart of the basic steps in singing transcription.

The scope of this chapter is restricted to the transcription of monophonic

singing performances in which only one note is sounding at a time. The pres-
ence of other instruments is not allowed. The singing style is not restricted
but singing with lyrics, syllables, or humming is considered, and the 'quality'
of the performance or the recorded signal is not limited in any way. It should
be noticed that we do not try to transcribe lyrics but only the note pitch
labels and their non-quantized timing information.
There are some research areas that are closely related to singing transcrip-
tion. These include the analysis, synthesis, and transformations of singing
voice [44], [337], the detection of singing-voice segments in polyphonic music
[33], melody extraction from polyphonic music [435], [228], and music infor-
mation retrieval (MIR) systems in general [455].
Currently, some transcription software are available on the Internet which
are more or less suitable also for singing signals. These include the MAMI C+-|-
library^ for singing transcription [136] and commercial applications such as
Autoscore, Digital Ear, Solo Explorer, and Akoff Music Composer. Free soft-
ware demonstrations can be downloaded to give an idea of the current capa-
bilities of software-based singing transcription. Also, there exist auto-tuning
applications which correct inaccuracies in singing pitch instead of producing
actual transcriptions. These are widely used in professional music production.

12.1.2 Organization of This Chapter

This chapter is organized as follows. Section 12.2 introduces singing-voice

production by humans and considers the relationship between singing sounds
and the corresponding musical notations. The rest of the chapter follows the
basic steps of singing transcription systems shown in Fig. 12.1.
Section 12.3 discusses the pre-processing of singing signals and the mea-
surement of acoustic features (such as the FO) from them. The pre-processing
aims at making the feature extraction more robust against additive noise and
the formant structure of singing sounds. Section 12.4 considers the conversion
of acoustic features into a sequence of notes and rests. This involves the seg-
mentation and labelling of notes, for which there exist two main approaches:
the segmentation and labelling can be performed either jointly or as separate

^Available at http: //www. ipem. ugent. be/MAMI

364 Matti Ryynanen

steps in cascade. Some systems also apply pre-trained acoustic or musicolog-

ical models. The resulting note sequence may be exposed to some additional
post-processing to improve the transcription quality. Section 12.5 concludes
the chapter.

12.2 Singing Signals

The human vocal organ produces several types of voice sounds, such as speech,
laughing, or whispering, in addition to singing. In contrast with the other
voice sounds, singing exhibits musically meaningful information, and here we
assume that singing sounds can be associated with notes in musical transcrip-
tions.
In this section, we study the act of singing in order to understand the
target signals of our transcription systems. First, we consider how the vocal
organ produces singing sounds and controls the acoustical properties of them.
Second, we discuss how humans associate singing sounds with notes and what
are the problematic properties of singing signals from the transcription point
of view. For an extensive study on the singing voice and its properties, the
reader is referred to the work of Sundberg in [613], [614].
The following discussion concentrates on voiced sounds, i.e., sounds with a
distinguishable pitch, although singing sounds can be considered as
more or less modified speech sounds and include also unvoiced and transient-
type sounds. These affect the note segmentation in singing. For a more exten-
sive review of speech sounds, please refer to [536], [321].

12.2.1 Production of Singing Sounds

Singing sounds are produced by the human vocal organ, which consists of
three basic units: (i) the respiratory system, (ii) the vocal folds, and (iii) the
vocal tract [614]. The sound production process is as follows. First, the respi-
ratory system creates an overpressure of air in the lungs, called the subglottic
pressure, which results in an air flow through the vocal folds. The vocal folds
start to vibrate and chop the air flow into a sequence of quasi-periodic air
pulses, thus producing a sound with a measurable fundamental frequency.
The sequence of air pulses is called the voice source and the process of sound
generation via the vocal fold vibration is referred to as phonation. At the fi-
nal stage, the voice source passes through the vocal tract, which modifies the
spectral shape and determines the timbre of the voice sound. This stage is
referred to as articulation and it controls the production of different speech
sounds and lyrics in singing.
Figure 12.2 shows a block diagram of the singing-sound production process.
The vocal-organ units control various acoustic properties of singing sounds,
including their fundamental frequency, timbre, and loudness. These are dis-
cussed in more detail in the following.
 12 Singing Transcription 365
ACTIVITY: BREATHING PHONATION ARTICULATION

SUBGLOTTIC VOICE VOICE

VOCAL RESPIRATORY PRESSURE SOURCE SOUND
ORGAN VOCAL FOLDS VOCAL TRACT
SYSTEM
UNIT:
1 r ^ , ' 1
'' 1
CONTROLLED i
VOICE 1 TYPE OF FUNDAMENTAL FREOUENCY ,' FORMANT
SOUND 1 PHONATION . ^ ^ FREQUENCIES
PROPERTIES:

LOUDNESS VOICE TIMBRE

Fig. 12.2. The stages of the singing-sound production. The dashed arrows indicate
the acoustic properties controlled by the three vocal organ units. (Modified from
Sundberg [613, p. 10], used by permission.)

Fundamental frequency

The fundamental frequency of a singing sound is primarily controlled by the

vocal folds. The term phonation frequency refers to the vibration frequency of
the vocal folds and, during singing sounds, this is the fundamental frequency
of the generated tone. The vibration is mainly controlled with the muscula-
ture of the vocal folds and to some extent with the amount of the subglottic
pressure so that the greater the subglottic pressure, the higher the phona-
tion frequency. In addition to the fundamental frequency, the generated tone
includes frequency components called the overtone partials at the integer mul-
tiples of the fundamental frequency. In an idealized case, the amplitudes of
the overtone partials can be expected to decrease by about 12 dB per octave
[613, p. 64].
The phonation frequencies may range from around 100 Hz for male singers
to over 1 kHz for female singers. The term register refers to a range of
phonation frequencies where the singing sounds are produced in a similar
manner, thus producing similar timbral characteristics. Basically, there ex-
ist two registers for male voices (normal and falsetto) and three registers for
female voices (chest, middle, and head).

Timbre

The vocal tract acts as the most important controller of the singing sound
timbre at the articulation stage. It functions as a resonating filter which em-
phasizes certain frequencies called the formant frequencies. These depend on
the configuration of the articulators, including the jaw, the tongue, and the
lips. In voiced sounds, the two lowest formants contribute to the identification
of the vowel and the higher formants to the personal voice timbre.
In addition, singing sound timbre is affected by the amount of subglot-
tic pressure and the tension of the vocal folds, resulting in different types
of phonation. The phonation types include pressed, normal, flow, breathy,
and whisper phonation, given in decreasing order of subglottic pressure. In
366 Matti Ryynanen

the normal phonation, the vocal folds are not completely closed. When the
subglottic pressure is high, the vocal folds close more rapidly and produce a
'pressed' sound. In the breathy and whisper phonations, the amount of sub-
glottic pressure is insufficient to properly vibrate the vocal folds and a part of
the air flow remains unchopped, thus producing a breathy-sounding phona-
tion. In the flow phonation, the produced sound is neither pressed nor leaky,
which is ideal for singing.

Loudness

The loudness of singing is influenced at several stages of voice production.

The loudness is mainly controlled with the amount of subglottic pressure and
the type of phonation. Although the subglottic pressure is smaller in the flow
phonation than in the normal phonation, the sound pressure level of the voice
source is maximal in the flow phonation.
Trained singers can modify their vocal tract configuration so that the
phonation frequency and a formant frequency match in order to create a
louder sound. This is known as the singer^s formant. In the both extremes
of the register, the control over loudness is more restricted, especially for
untrained singers.

12.2.2 Singing Sounds and Notes

The phonation frequency enables the association of a singing sound with a

musical note. If a singer could perform each note with a stable phonation fre-
quency corresponding to the note pitch, the singing transcription task would
be readily accomplished. However this is not the case, because singing perfor-
mances possess both intentional and unintentional deviations from the nom-
inal note pitches. Deliberate phonation-frequency deviations are commonly
used to enhance the expressiveness of singing performances, whereas the un-
intended deviations are mostly due to the lack of voice training.
Figure 12.3 shows the fundamental frequency curve and the loudness
curve of the note A4 performed by a female singer. Listening to the note
performance, one unambiguously associates a note with A4, although the
phonation-frequency curve is not stable during the note. To address this per-
ceptual association of phonation frequencies with notes, the following discus-
sion considers issues such as vibrato, tremolo, glissando, legato, and out of tune
singing.

Vibrato and tremolo

The term vibrato refers to the modulation of the phonation frequency during
a performed note. Vibrato can be characterized with the rate and the depth of
the modulation. The rate typically varies between 4-7 Hz [425], and the depth
between 0.3-1 semitones. In [530], the depth of vibrato was measured in ten
recordings performed by professional singers and was reported to range from
 12 Singing Transcription 367

480 A)FUNDAMENTAL FREQUENCY

> 460 GLISSANDO _ NOTE:
VIBRATO
^ 440
m 420
O 400
S 380
LL

TIME (s)

Fig. 12.3. A) The fundamental frequency curve and B) the loudness curve mea-
sured from a note A4 (440 Hz) performed by a female singer. The FO curve clearly
shows glissando and vibrato, whereas the loudness curve shows tremolo. Notice the
correlation between these two.

0.34 to 1.23 semitones. Experimental results with h u m a n listeners indicate

t h a t the mean of the fundamental frequency during a vibrato note is close to
the perceived pitch of the note [589].
T h e fluctuation of loudness is usually referred to as tremolo. In Fig. 12.3,
t h e phonation-frequency and loudness curves imply vibrato and tremolo effect
during the performed note, respectively. T h e vibrato rate is approximately
6.5 Hz and the vibrato depth varies from half a semitone to slightly over
one semitone. T h e figure also shows a high correlation between vibrato and
tremolo.

Glissando and legato

T h e t e r m glissando refers to a phonation frequency slide to the note pitch.

Glissando is usually employed at the beginning of long notes: these often
begin flat (too low) and t h e phonation frequency is matched t o the note pitch
during the first 200 ms of a note [570]. Figure 12.3 shows a clear glissando in
which the note begins over two semitones flat. Singing in legato means t h a t
consecutive notes are tied together, for example by using glissandi to change
the note pitch.

Singing out of tune

Singing out of t u n e refers to a situation where a note pitch differs annoy-

ingly from the tuning of the other notes within a performance. Accord-
ing to the listening test results reported by Sundberg in [614], the mean
phonation frequency may deviate 0 . 0 7 semitones from the nominal note
pitch for notes which were generally judged to be in tune. However, mean
phonation-frequency deviation larger t h a n 0 . 2 semitones can be acceptable
368 Matti Ryynanen

at unstressed metrical positions, during tragic song mood, or when the phona-
tion frequency is sharp (too high) rather than flat. It was also observed that
the direction of the deviation was related to the musical context [614].
Other types of tuning problems arise when singing is performed without
an accompaniment. In general, singers are not able to perform in absolute
tuning and, moreover, the tuning may drift, meaning that a singer gradually
changes the baseline tuning over time.

12.3 Feature Extraction

The first step in singing transcription is to measure acoustic cues that enable
the detection of notes in singing signals. The acoustic cues are here referred to
as features, and the term feature extraction refers to the measurement process.
The most salient features include the fundamental frequency and features re-
lated to note segmentation. Features are commonly extracted frame-by-frame
from acoustic waveforms and a pre-processing stage may precede this to fa-
cilitate the extraction process.

12.3.1 Pre-Processing

Pre-processing of the input signals aims at facilitating the feature extraction

process. If singing has been recorded in a noisy environment or with low-
quality equipment, the signal may include additive background noise which
needs to be attenuated. The noise reduction problem has been studied for
decades in speech processing and, to some extent, the same methods can be
applied in singing transcription. A comprehensive study of noise reduction can
be found in [638]. An issue related to noise reduction is audio restoration that
aims at improving the quality of old music recordings, for which an interested
reader is referred to [214].
The two main steps in noise reduction are the estimation of the noise com-
ponent in the signal and its subsequent suppression. For example, PoUastri
estimated the background noise spectrum during the time segments where
singing activity was not detected [528]. The noise component was then sup-
pressed using linear power-spectral subtraction.
In addition to noise reduction, spectral whitening can be applied at the
pre-processing stage to flatten strong formants in the signal spectrum. This
aims at facilitating the measurement of fundamental frequency, since it is most
salient in the voice source signal before the vocal tract influence. In [627],
inverse filtering with a warped linear prediction (WLP) model was applied
to perform spectral whitening. The idea of WLP is the same as for ordinary
linear prediction, but implemented on a frequency scale resembling that of
human hearing.
 12 Singing Transcription 369

12.3.2 Fundamental Frequency Estimation

Fundamental frequency is the most important feature in singing transcription
systems. FO estimation has been widely studied in speech processing, and most
of the state-of-the-art singing transcription systems use algorithms originally
designed for speech signals. The following survey introduces a few useful FO
estimation algorithms for singing transcription. For a more thorough review
of different methods, see [289], [290], [618], [134].

Time-Domain Methods
Autocorrelation-based FO estimators are widely used in singing transcription.
The idea of these is to measure the amount of correlation between a time-
domain signal and its time-shifted version. For periodic signals, the autocor-
relation function has local maxima at time shifts that equal the fundamental
period and its multiples.
Given a sampled time-domain signal s{k) and a frame length VF, the short-
time autocorrelation function rt{r) at time t is defined as
t+w-i
rt{T)= Yl s{k)s{k^r), (12.2)
k=t

where r is called the lag. Figure 12.4 shows rt{T) as calculated for a frame
of a singing signal. The function rt{r) peaks at lags which correspond to
the multiples of the fundamental period. A fundamental frequency estimate is
obtained by dividing the sampling rate of the signal with the smallest non-zero
lag value for which rt{r) reaches a value above a chosen threshold.
The autocorrelation method is straightforward to implement and easy to
use. A drawback is, however, that the method is sensitive to formants in
signal spectrum and therefore tends to make octave errors. Spectral whitening
makes the autocorrelation more robust in this respect. Autocorrelation has
been employed in the singing transcription systems of Ghias et al. [207], Shih
et al. [587], [586], [588], and Wang et al. [659], for example.
YIN algorithm
The YIN algorithm for FO estimation was proposed by de Cheveigne and
Kawahara [135]. It resembles the idea of autocorrelation but introduces cer-
tain improvements that make it more convenient to use. The YIN algorithm
was successfully used for singing transcription by Viitaniemi et al. [643] and
Ryynanen and Klapuri [558]. Given that s{k) is a discrete time-domain signal
with sampling rate /s, the YIN algorithm produces a FO estimate as follows.
1. Calculate the squared difference function dtir) where r is the lag:
t-\-w-i
dt{T)= Yl {s{k)-s{k + T)f. (12.3)
k=t
370 Matti Ryynanen

150 200 250 300 350 450 500

samples

Fig. 12.4. A) An excerpt of a singing signal with sampling rate 44.1 kHz. B) The
ACF rtir) calculated using (12.2). C) The cumulative-mean-normalized difference
function d^r) calculated using (12.4) (YIN algorithm). See text for details.

2. Derive the cumulative-mean-normalized difference function d[{T) from

dtir):

'^'^^^^ = I dtir) I [(1/r) E ; = I dm , otherwise. (^2.4)

3. Search for the smallest r for which a local minimum of dl^ij) is smaller
than a given absolute threshold value K. If no such value can be found,
search for the global minimum of G?J(r) instead. Denote the found lag value
with T' .
4. Interpolate the d\{T) function values at abscissas {r' \,r',T' + 1} with
a second-order polynomial.
5. Search for the minimum of the polynomial in the continuous range {r'
l , r ' + 1) and denote the corresponding lag value with f. The estimated
fundamental frequency is then / S / T .
The normalization in (12.4) makes dj(r) independent of the absolute signal
level and removes a spurious local minimum around the lag zero, thus making
the subsequent thresholding in Step 3 more convenient to implement. The
threshold value K = 0.15 was used in [558]. The interpolation in Step 4 in-
creases the numerical accuracy of the FO estimate.
Panel C in Fig. 12.4 shows the function o?J(r) calculated from the signal
in panel A, and the dashed line indicates the threshold value K 0.15. The
first local minimum below the threshold is found at r' 92 and then refined
to f = 92.37 by the interpolation. For sampling rate /s 44.1 kHz, this
yields a FO estimate /g/f ?^ 477 Hz. Compared to the ACF in panel B, Vtir)
and d!^{T) exhibit maxima and minima at the same lag values. To conclude
 12 Singing Transcription 371
PREDICTED NEAREST MEASURED
HARMONICS MATCH PARTIALS

4F0 -< *"

3F0 - . ^ ^ PREDICTED-TO-MEASURED
"" - ~ *^
2F0 -*- - V"-"-"-V_VJt ^ MEASURED-TO-PREDICTED

TRIAL FO ^

Fig. 12.5. The two-way mismatch error calculation is baised on combining the
predicted-to-measured and mea^ured-to-predicted errors. The two-way matching
makes the method robust for both 'too low' and 'too high' octave errors. (After
Maher and Beauchamp [428].)

with, normalization and thresholding are important in autocorrelation-based

methods in general, and the YIN algorithm proposes one useful alternative
for this.

Frequency-Domain Methods

Cepstrum-based FO estimators originated in speech processing (see Chapter 2

for the definition of the cepstrum). The fundamental period of a signal can
be found from the cepstrum in a way similar to the ACF-based methods.
Cepstrum-based FO estimators perform well for speech sounds with strong
formant structures, and they can also be used for singing transcription. How-
ever, these FO estimators are rather sensitive to noise, since the logarithm
operation involved in the cepstrum calculation raises the level of the noise
floor in relation to harmonic components in the spectrum. A cepstrum-based
FO estimator was used in the query-by-humming system of Liu et al. [411],
for example.
Maher and Beauchamp proposed a FO estimation method for music signals
which was based on a two-way mismatch procedure [428]. The method searches
for the most likely fundamental frequency among a set of trial FO values.
The amplitudes and the frequencies of prominent sinusoids are first measured
from the short-time spectrum of a signal frame. Given a trial FO, the measured
partials are compared to the predicted harmonics of the trial FO (measured-to-
predicted error) and vice versa (predicted-to-measured error), thus producing
two error measures for the FO estimate in question. Figure 12.5 illustrates
the matching process. The two error measures are then combined to derive
a total error measure. The trial FO with the smallest total error is considered
as the best FO estimate within the frame. Details of the algorithm can be
found in [428]. The algorithm is relatively straightforward to implement, and
it has been used to separate duet signals [427] and to analyse the melody line
in polyphonic recordings [519].
Peak-picking FO-estimation algorithms search for peaks in the magnitude
spectrum and group them to derive an FO estimate. In the systems of Pollastri
372 Matti Ryynanen

AUDIO COCHLEAR FREQUENCY FOs

BPF AMPEX
FILE PROCESSOR SPLITTER (HPF) STRENGTHS

NEURAL FIRING PATTERNS

Fig. 12.6. A block diagram of the auditory model-based FO estimator used by

Clarisse et al. [92].

[528] and Haus and Pollastri [278], an FO estimate was determined by searching
for the most prominent peak in a spectrum between 80 Hz and 800 Hz for
which at least two clear harmonics were found. Another peak-detection scheme
was applied by Zhu and Kankanhalli in [689], where a set of the highest peaks
in the spectrum were detected and a rule-based peak-grouping algorithm was
employed to obtain an FO estimate.

Auditory Model-Based Methods

Clarisse et al. used an auditory model-based FO estimator in their singing tran-

scription system [92]. Figure 12.6 shows the block diagram of the method. The
algorithm was originally designed for speech signals and its detailed explana-
tion can be found in [637]. In brief, the algorithm works as follows.
An audio file is first bandpass filtered to model the frequency response of
the outer and the middle ear. The signal is then fed into a cochlear proces-
sor, where a filter bank models the frequency selectivity of the inner ear. The
signal at each band is processed by a hair-cell model in which the signal is
half-wave rectified (i.e., negative signal-sample values are set to zero) and
its dynamic range is compressed. The output of each hair-cell model is fol-
lowed by a lowpass filter which extracts the envelope of the signal in the
channel.
The envelope signals in different channels are processed by a pitch extrac-
tion module called AMPEX. It performs pseudo-autocorrelation analysis and
sums the results across channels to obtain a global pseudo-autocorrelation
function. The major peaks in the function are searched for to obtain a set of
pitch candidates and their corresponding strengths. The final pitch estimate
is chosen by analysing the continuity of the pitch estimates in the analysis
frame and its surrounding frames. If the strength of the final pitch estimate
exceeds a threshold value, the frame is considered to be voiced. Later, De Mul-
der et al. concluded that the AMPEX algorithm was unsuitable for the FO
estimation of very high-pitched sounds and proposed another FO estimation
algorithm working in parallel with AMPEX [137].
A drawback of the auditory model-based algorithms is that they are
computationally quite demanding. As an exception to this, Tolonen and Kar-
jalainen have proposed an auditory model-based FO estimator which is com-
putationally efficient [627]. The method is described in Chapter 8 but, to our
knowledge, it has not yet been applied to singing transcription.
 12 Singing Transcription 373

12.3.3 Features for Note Segmentation

Singing transcription requires the temporal segmentation of notes, i.e., the

determination of where notes begin and end in time. This is usually done
based on two types of features: those that indicate silence, noisy, or voiced
segments in audio signals; and those that imply the beginning or the ending of
a note. The first type of features include signal energy, the degree of voicing,
and zero crossing rate, and the second type of features include phenomenal
and metrical accents (to be explained below).

Energy

Features related to the signal energy are widely used for note segmentation,
with the assumption that the signal level reflects the loudness of the target
singing voice. Segments where the signal energy exceeds a given threshold
value are often considered to be notes and the other segments are treated as
silence or background noise. Energy-related measures are straightforward to
use and work well if the notes are separated with more quiet regions. However,
this is not the case in most signals; usually there are also legato-type note
transitions. Therefore, robust note segmentation cannot be based on energy
measures alone.
One of the most often used energy measures is the root-mean-square
(RMS) energy (see Chapter 6 for the RMS definition). This feature has been
apphed by Haus and Pollastri [278], Pollastri [528], McNab et al. [454], [455],
and Shih et al. [587], [586], [588], for example. Different variants of energy
calculations have been applied in note segmentation, including the systems
by Clarisse et al. [92], Liu et al. [411], and Orio and Sette [489].

Voicing

A more reliable feature for note segmentation is the degree of voicing of a sig-
nal frame. Voiced frames possess clear periodicity, whereas unvoiced frames
can represent transient noise with a great amount of signal energy, or just
silence. Commonly, the voicing determination is embedded within the FO es-
timation algorithms. In autocorrelation-based FO estimators, for example, the
degree of voicing is straightforwardly given by the ACF function value at the
lag corresponds to the estimated fundamental period, divided by the value at
lag zero. If the ratio of these two does not exceed a given threshold, the frame
is considered to be unvoiced.
When the YIN algorithm is used for FO estimation, the degree of voic-
ing can be directly derived from the value d[{f) of the cumulative-mean-
normalized difference function defined in (12.4). The d[{T) value itself de-
scribes the amount of non-periodicity in FO measurement. To obtain a voicing
value, we have to map the d[{T) value to a voicing feature for example as

u{f) = 1 - < ( f ) . (12.5)

374 Matti Ryynanen

The voiced/unvoiced decision can also be based on the zero crossing rate
(ZCR) together with an energy measure, as done in [528], [421]. ZCR loosely
describes the brightness of the sound. High ZCR values imply transient or
noisy segments since these tend to have lots of energy at high frequencies.
ZCR is defined in Chapter 6.

Accents

Features indicating note beginnings include the degree of phenomenal accent

and the degree of metrical accent as a function of time. Phenomenal accents
refer to moments that have perceptual emphasis in music signals, whereas
the metrical accent corresponds to the underlying pulse of a music perfor-
mance [404].
Klapuri proposed a method for estimating both of these accents in music
signals [349]. In his system, a (phenomenal) accent signal is measured based
on the amount of spectral change as a function of time. Briefly, the signal
power is first measured at 36 sub-bands. At each band, the power of the sig-
nal is computed and then log-compressed and smoothed over time by lowpass
filtering. The power envelopes are then subjected to time differentiation and
half-wave rectification to measure changes in intensity. Finally, the resulting
signals are summed across the bands to produce the accent signal. The met-
rical accent signal is constructed by inferring regularities in the accent signal
and by estimating the temporal pulse of the music performance. The method
of Klapuri can be used as a tempo estimator and is capable of following the
tempo changes during music performances. For further discussion on beat-
tracking and musical metre analysis, see Chapter 4.
Ryynanen and Klapuri used the accent signal and the metrical accent
signal as features in a singing transcription system [558]. The accent signal
indicated the singing note beginnings whereas the metrical accent signal was
used to predict possible note beginnings according to the estimated tempo of
the performance.

12.3.4 Summary of the Features and an Example

Table 12.1 summarizes the features discussed in this section. Figure 12.7 shows
a selection of features extracted from a short singing excerpt containing five
notes performed by a professional female singer. From top to bottom, the pan-
els in the figure show the recorded singing waveform, fundamental frequency
estimates, the degree of voicing, RMS energy, the accent signal, and the zero
crossing rate. The panel with FO estimates also shows a manual transcription
of the notes in the performance (B4, A4, A4, Ftt4, and A4). The FO estimates
and the degree of voicing were obtained using the YIN algorithm and (12.5).
At the beginning of the excerpt there is silence, which can be observed
from the RMS and the degree of voicing. When the first note begins, the
accent signal has a clear peak. Interestingly, the second note begins with an
 12 Singing Transcription 375

Table 12.1. Acoustic features for singing transcription.

Feature Usage Methods Comments
FO note labelling autocorrelation well-motivated & understood
YIN related to the ACF
cepstrum robust against formants, sen-
sitive to noise
two-way mismatch rather robust, intuitive
auditory model-based computationally complex
Energy segmentation RMS -h variants reflects the loudness
Voicing segmentation ACF value side-product of FO estimation
ZCR reflects the brightness
Accents segmentation (phenomenal) accent indicates note onsets
metrical accent predicts note onsets

almost two-semitones fiat glissando. Simply by following the FO estimates, the

beginning of this note would be easily interpreted as note G4 {^ 390 Hz). In
a manual transcription, however, the choice of a single note A4 is obvious.
The last note includes strong vibrato and tremolo as well as a glissando in the
beginning of the note. Although the accent signal indicates note beginnings
quite reliably, especially strong vibrato causes false peaks to the accent signal,
as can be seen during the last note. The zero crossing rate has a peak at time
2.8 seconds caused by the consonant / s / .
As the example shows, it is not always obvious what notes should be
transcribed when given the extracted features, and this makes the singing
transcription task rather difficult. Conversion of the frame-level features into
note pitch labels is considered in the next section.

12.4 Converting Features into N o t e Sequences

The greatest challenge in singing transcription lies in the conversion of frame-
level features into note sequences. Singing performances exhibit inaccuracies
both in pitch and timing, usually as a consequence of expressive singing or
the lack of voice training. Therefore, a singing transcription system may not
assume that the notes are performed in an absolute tuning, that the phonation
frequency is stable during the notes, or that the notes are performed in a
rhythmically accurate manner.
The features-to-notes conversion involves note segmentation and labelling.
These two steps can be performed either (i) in a cascade or (ii) jointly.
The former approach is here referred to as the segment-and-label approach
where singing recordings are first segmented into notes and rests, and each
note segment is then assigned a pitch label. The joint segmentation and la-
belling of notes usually applies statistical models of note events and is here
376 Matti Ryynanen

2 0.5

^ jf\f"^^f^S/^'

" % /

1/
z
o
5 0.5
.5||/'

z
LU
o
o
<

0 0.5 1.5 2 2.5 3.5

TIME (s)

Fig. 12.7. Selected features extracted from an excerpt performed by a professional

female singer. The panels show (in top-to-bottom order) the singing recording wave-
form FO estimates, the degree of voicing, RMS energy, the accent signal, and the
zero crossing rate. The excerpt was manually transcribed to contain five notes, B4,
A4, A4, FtJ4, and A4, and they are shown in the panel of fundamental frequencies.
See text for details.

referred to as the statistical modelling approach. Both of these approaches may

apply musical rules or musical context to obtain a better transcription result.
At an optional post-processing stage, the transcription can be enhanced, for
example by adding a parametric representation of an estimated expression to
the transcription.
 12 Singing Transcription 377

RMS
0.2 h - - ONSET THR
OFFSET THR
I I NOTE SEGMENTS
0.1

4 5 6
TIME (s)

Fig. 12.8. Note segmentation based on RMS-energy thresholding.

12.4.1 Segment-and-Label Approach

As mentioned above, the idea of the segment-and-label approach is to first

segment an audio signal into notes and rests and then to assign a pitch label to
each note segment. This approach has been taken in the singing transcription
systems of McNab et al. [454], [455], Haus and Pollastri [278], Pollastri [528],
Lu et al. [421], Wang et al. [659], and Clarisse et al. [92].

Note Segmentation

Note segmentation is usually based on features related to the signal energy or

voicing, or on abrupt changes in fundamental frequency. The methods usually
apply fixed or adaptive thresholds to decide whether a note boundary has
occurred or not.
McNab et al. used the RMS energy for note segmentation [454]. When the
signal energy exceeded a given threshold, the method interpreted this as a note
onset, and similarly, when the signal energy dropped below another threshold,
a note offset was inferred. In addition, the segment had to be at least 100 ms
long to be accepted. Figure 12.8 shows the measured RMS contour of ten
singing notes performed with a / d a / syllable. The onset and offset thresholds
are set to 50% and 30% of the overall RMS energy as in [454]. The note offsets
are accurately found, whereas for example the first note onset is detected at
1.1s although it should be at 0.5 s.
As an alternative for the energy-based note segmentation, McNab et al.
used a scheme where slowly varying regions of FO estimates were grouped. A
note boundary was detected when an FO estimate deviated more than half a
semitone from the average of the previous estimates during the segment.
Pollastri used both the signal energy and FO changes for note segmentation,
in addition to ZCR, which was used to make voiced/unvoiced decisions [528].
Clarisse et al. performed note segmentation based on adaptive adjustment of
threshold values for signal energy. Later, the system of Clarisse et al. was
improved by adding a number of decision rules for note segmentation, so as
to handle notes with legato, vibrato, and tremolo [137].
378 Matti Ryynanen

N o t e Labelling

Note labelling follows the note segmentation. At this stage, each note segment
is assigned a pitch label such as an integer MIDI note number or a note name.
The most important question here is how to determine a single label for a note
segment where FO estimates are widely varying and possibly out of absolute
tuning. The different note-labelling schemes differ from each other mainly
in terms of how they handle the singer's tuning. This has been addressed
using three different assumptions:
1. The singer performs in absolute tuning (note A4 corresponds to 440 Hz);
2. The singer has a relative tuning in mind and makes consistent deviations
from the absolute tuning;
3. The singer allows the baseline tuning to drift during a performance.
Corresponding to the three assumptions, the transcription systems either per-
form no tuning, estimate a constant tuning, or perform time-adaptive tuning,
respectively. Three systems applying these three approaches are briefly intro-
duced in the following.

No tuning

The system of Clarisse et al. aimed at transcribing notes as precisely as pos-

sible according to singing pitch instead of trying to infer the intention of the
singer [92]. The FO for a note segment was calculated as the arithmetic mean
of the FO estimates (Hz) in the central part of the note segment. The FOs
which differed more than 10% from the mean were discarded and the mean
was recalculated. The note pitch label was then obtained by rounding the
resulting mean FO into an integer MIDI note number.

Constant tuning

Haus and Pollastri assumed that the performed notes differ a constant amount
in semitones from the absolute tuning [278]. Figure 12.9 shows a block diagram
of their pitch-labelling process. Given a note segment, the FO estimates within
the segment were first 3-point median filtered to remove FO outliers. Then a
group of four contiguous frames with similar FO values constituted a block.
At the block level, legato with note pitch change was detected when pitch
between adjacent blocks had changed more than 0.8 semitones. In the case
of a detected legato, the note segment was divided into two new segments.
Otherwise, the adjacent blocks constituted a note event for which the FO was
calculated as the arithmetic mean of the FO in the blocks and represented as
an unrounded MIDI note number. This process was repeated for each note
segment, resulting in note segments with unrounded MIDI note labels.
To determine the deviation from the absolute tuning, the authors calcu-
lated a histogram of distances from the unrounded MIDI note numbers to their
nearest integer numbers. The highest peak in the histogram then indicated the
 12 Singing Transcription 379

NOTE ^ FRAME FO REMOVE FO GROUP

SEGMENTS ESTIMATION OUTLIERS FRAMES

EVENT(S) 1 OC
^^
NOTE ^ CONSTANT AVERAGE IF LEGATO,
SEQUENCE TUNING PITCH SPLIT SEGMENT
UNROUNDED
MIDI NOTES

Fig. 12.9. Pitch-labelling process of Haus and Pollastri [278].

most-often occurring offset from the absolute tuning and, subsequently, every
note segment label was shifted by this offset, thus minimizing the rounding
error. Finally, the shifted MIDI note labels were rounded to integers, thus
obtaining the note pitch labels.

Time-adaptive tuning

McNab et al. considered a labelling technique which continuously adapted to

the singer's tuning during a performance [454]. For the first note segment, a
histogram of the FO estimates in the segment was calculated and the most
prominent peak in the histogram was rounded to a MIDI note number. For the
next note segment, however, the histogram was shifted by the amount of the
rounding done in labelling the previous note. In other words, the tuning was
constantly adjusted according to the rounding errors made while labelling the
preceding notes. This approach, however, tends to adjust tuning too much at
a time. In the worst case, the maximum rounding error in each note labelling
can be half a semitone. Thus the reference tuning might have changed one
octave (twelve semitones) after labelling 24 notes.

12.4.2 Statistical Modelling of Notes

The statistical approach discussed in this section applies pre-trained acoustic

models of note events and performs the note segmentation and labelling
jointly. This has certain advantages in the singing transcription problem in
particular. First, it is theoretically advantageous to optimize the note segments
and their labels jointly instead of doing these as two consecutive steps. Sec-
ondly, statistical models allow representing uncertain information and learning
from examples, thus providing a more convenient way to deal with the highly
varying acoustic data.
Statistical note models are derived by measuring low-level acoustic features
during note events and by training a parametric statistical model to describe
the behaviour of the features during the events. The models can then be used
to calculate the likelihoods of different notes in previously unseen material.
The note sequence which maximizes the overall posterior probability according
to the note models is considered to be the transcribed note sequence.
380 Matti Ryynanen

Hidden Markov models

In practice, all the current singing transcription systems applying statistical

note modelling have been based on hidden Markov models (HMMs). HMMs
are widely used to model phonemes and words in automatic speech recognition
and, therefore, the theory of HMMs is well established. The following presen-
tation introduces the basic idea of HMMs and their usage in note modelling.
For a more formal consideration of HMMs, see Chapter 2. A good tutorial on
HMMs can be found for example in [534].
A hidden Markov model is a state machine which models stochastic signal
sources. The HMM is mathematically formulated by the following parameters:
1. A set of states = {^1,62,... ,e;} within the model where E is the
number of states.
2. State-transition probabilities P{9t = ej \6t-i e^), i.e., the conditional
probabilities that state ei is followed by state ej within a random state
sequence Oi-t- In particular, a direct transition from e^ to ej is not possible
ifP(^t = e , | ^ t - i - e O = 0.
3. The observation likelihood distributions p{xt \0t = ej), i.e., the likelihoods
that an observation vector x is emitted from state ej G 0. Here the
observation vector consists of the acoustic features extracted in frame t.
4. The initial probabilities P{6i = ei) and the final state probabilities
P{6T = ^i), i-e., the probabilities that a state sequence begins or ends
to state ei.
The state-transition probabilities, the observation likelihood distributions,
and the initial and the final state probabilities can be learned from acoustic
data either with supervised or unsupervised training techniques, depending on
whether the underlying state sequence 6i:t is known a priori or not. In the case
of singing transcription, the training data consists of the frame-level features
extracted from singing recordings, and the training is usually performed in an
unsupervised manner with the Baum-Welch algorithm, which is described for
example in [534].
Once the HMM parameters have been learned, the state sequence which
maximizes the posterior probability of the observed data can be estimated
using the Viterbi algorithm [654]. See Chapter 2 for a description of the
Viterbi algorithm. An alternative state-sequence estimation scheme is the
token-passing algorithm, which is designed for finding the most probable path
through a network of connected HMMs. For the algorithm details, see [685].
Briefly, the token-passing algorithm propagates tokens which represent alter-
native paths through the network. Inside a HMM, each state contributes to
the weight of the tokens by the observation likelihoods and the state-transition
probabilities. When a token is emitted out of an HMM, the algorithm iden-
tifies the boundary between the connected HMMs and appends it to a list of
boundaries for the purpose of backtracking the most probable path after the
 12 Singing Transcription 381

analysis. The tokens are further weighted with predefined transition probabil-
ities between different HMMs. Eventually, the most probable path is defined
by the token with the maximum posterior probability and the corresponding
boundaries between different HMMs in the network.

Note-Event Model of Ryynanen and Klapuri

Ryynanen and Klapuri proposed a singing transcription system based on mod-

elling note events with HMMs [557], [558]. The system performs note segmen-
tation and labelling simultaneously and also utilizes the musical context in
singing performances. The note-event model and its usage are described in
the following, whereas the use of musical context is considered later in Sec-
tion 12.4.3.

Note-event model

Note events are described with a three-state left-to-right hidden Markov model
where P{6t = Cj \ 9t-i = e^) 7^ 0 only when j = z or j = i -h 1. The state ci in
the model represents the typical acoustic characteristics of the ith temporal
segment of a performed singing note. The model uses three features: funda-
mental frequency estimates (represented as unrounded MIDI note numbers),
the degree of voicing, and the accent signal. The features are extracted as ex-
plained in Section 12.3. Different notes are represented with a separate HMM
for each MIDI note n = 3 6 , . . . , 79. For note n, the features in frame t form
the observation vector Xt where the difference between the fundamental fre-
quency estimate and note n is used instead of the FO estimate directly. This
is referred to as the pitch difference AFQ:

AFo = Fo-n. (12.6)

The use of pitch difference facilitates the training of the model. Usually, there
is a limited amount of training data available, at least for each possible singing
note. Due the pitch-difference feature, it is possible to train only one set of
note HMM parameters with greater amount of data, and the same parameters
can be used to represent all the different MIDI notes.
The state-transition probabilities and the observation hkehhood distrib-
utions were estimated from an acoustic database containing audio material
performed by eleven non-professional singers. The singers were accompanied
by MIDI representations of the melodies which the singers heard through
headphones while performing. Only the performed melodies were recorded,
and later, the reference accompaniments were synchronized with the perfor-
mances. The reference notes were used to determine note boundaries in the
training material, and the Baum-Welch algorithm was then used to learn the
note HMM parameters.
Figure 12.10 illustrates the trained note HMM. The HMM states are shown
on top of the figure where the three states are referred to as the attack.
382 Matti Ryynanen

NOTE HMM

OBSERVATION LIKELIHOOD DISTRIBUTIONS IN EACH STATE:

PITCH
DIFFERENCE
(SEMITONES)

DEGREE OF
VOICING

0.02

ACCENT
0.01

Fig. 12.10. A note HMM with three states: an attack, a sustain, and a silence/noise
state. The arrows show the possible state-transitions. The observation HkeUhood
distributions are shown below each state for the features used (pitch difference, the
degree of voicing, and accent).
the sustain, and the silence/noise state, each of which corresponds to a time
segment within note events. The arrows show the possible transitions between
the states. The observation likelihood distributions shown below each state
express the typical behaviour of these features during the different segments
within note events. It is a little surprising that unsupervised learning leads to
such an intuitive interpretation of the three states of the note model:
Attack: The singing pitch may vary about a few semitones from the actual
note pitch. Since the mean of the pitch-difference distribution is at 0.5
semitones, this implies that notes usually begin slightly flat. Accent value
distribution is widely spread, which indicates the presence of large accent
values during note attacks.
Sustain: The variance of pitch difference is much smaller than in the attack
state. Most of the FO estimates stay within 0.2 semitone distance from
the nominal pitch of the note. In addition, the frames during the sustain
state are mostly voiced (i.e., the FO estimation can be reliably performed)
and the accent values are small.
 12 Singing Transcription 383

MUSICOLOGICAL MODEL
TOKENS _ ^ .

NOTE MODEL

'He-e^b^aaebdae^
-

CO
LJJ
h-
o
z
HaaeP/^aaep/daae
1 Y y
2 1 / v \
1 / X \

V J \
e-e^Haae
TIME

Fig. 12.11. Combination of the note models and the musicological modeL

Silence/noise: The FO estimates are almost random valued, since most of the
frames are unvoiced (notice the high likelihood of voicing values near 0.3).
The authors did not consider modelling the rests, since the note model
itself includes a state for silent or noisy segments. Thus the time instants
where the silent/noise state is visited can be considered as rests.

Using the note event model

The HMMs of different MIDI notes are joined into a note model network illus-
trated in Figure 12.11. The probability of a transition from one note to another
is determined by a musicological model (explained in Section 12.4.3). Notice
that the figure shows the note models at successive time instants although
there actually exists only one note model per MIDI note. Singing melodies
can be transcribed by finding the most probable path through the network
according to the probabilities given by the note HMMs and the musicological
model. The authors used the token-passing algorithm for this purpose.
The transcription system of Ryynanen and Klapuri was evaluated using 57
singing melodies from the same database that was used for training the note
models (the training signals were not included in the evaluation set). An error
rate below 10% was achieved when the transcriptions were compared to the
reference MIDI notes. Simply rounding FO estimates to MIDI note numbers
produced an error rate of about 20%. The use of note models reduced the
error rate to 13%, and including also the musicological model decreased the
error rate below 10%. A similar note model network was later applied by
the authors to the transcription of polyphonic music from various genres [559].
384 Matti Ryynanen

Other Systems with Note-Event Models

Orio and Sette used a HMM-based note modelling technique for transcribing
singing queries [489]. Their approach was similar to the system of Ryynanen
and Klapuri: each MIDI note was represented by its own HMM, which had
three states for modelling the attack, sustain, and rest segments in singing
notes. The used features included the logarithm of signal energy, spectral
energy on the first harmonics of each modelled note for FO detection, and
the first derivatives of these. The observation likelihood distributions were
derived by making statistical analysis of a set of labelled audio examples.
The HMMs of different notes were joined into a note network, and the Viterbi
algorithm was used to decide both the note segments and the pitch note labels
simultaneously, thus producing transcriptions of singing queries.
Their preliminary results indicated that most of the labelhng errors were
due to glissandi between notes. Another common error was to insert additional
note boundaries during note attacks. The authors discussed the interesting
possibility of using several attack states for modelling different types of note
beginnings and an enhanced sustain state with two additional states modelling
slight detunings upwards and downwards from the note pitch. However, these
ideas were not implemented in the reported system.
Viitaniemi et al. proposed an HMM for singing transcription, where the
HMM states corresponded to different notes [643]. The observation likelihood
distributions consisted of the FO value distributions for each note and the
state-transitions were controlled with a musicological model. The optimal
state sequence was found with the Viterbi algorithm in order produce note
labels at each time instant. In the optimal state sequence, transitions between
different states were interpreted as note boundaries. The system achieved error
rates around 13% for the database that was used by Ryynanen and Klapuri
in [558].
Shih et al. used a three-state HMM to model note events [587]. The features
included mel-frequency cepstral coefficients (MFCC), one energy measure, and
a pitch ratio. The pitch ratio was calculated as

log(Fo) - log(FJ^f) , (12.7)

where FQ (HZ) denotes a fundamental frequency estimate and FQ^^ a refer-

ence FO. Two different definitions for the reference FO were tested: FJ^^ was
defined as the arithmetic mean of the FO values (i) in the first detected note
segment or, alternatively, (ii) in the immediately preceding note segment. In
the former case, they trained 15 note models describing one octave of a major
scale upwards and downwards from the first detected note. In the latter case,
they trained models for one and two semitone intervals upwards and down-
wards with respect to the previous note. In addition, they defined a model for
segmenting the first note. Once the first note of a performance is segmented,
the mean FO is calculated as the reference FO. Depending on the selected set
of HMMs, the reference FO was then either kept fixed or updated after each
 12 Singing Transcription 385

detected note, leading to the constant and the time-adaptive tuning, respec-
tively (see Section 12.4.1). The time-adaptive tuning was reported to work
better for which approximately 80% of notes were correctly labelled.
Shih et al. soon modified their system to transcribe singing in two consec-
utive stages of note segmentation and labelling [588]. The note segmentation
was performed with two three-state HMMs, one of which modelled note seg-
ments and the other rests. The note and rest segments were then found using
the Viterbi algorithm. After a note segment was detected, it was labelled with
respect to its previous note by using pitch interval in semitones and the stan-
dard deviation of FO estimates within the note segment. These attribute values
were modelled with a Gaussian mixture model (GMM) and the interval label
was decided according to the most likely GMM. Some minor improvements
in transcription quality were achieved compared to their previous system. In
addition, the improved system operated in real time.

12.4.3 Utilizing Musical Context

So far, we have discussed singing transcription without any prior knowledge

about the musical relevance of individual notes or their relationships. Musical
context plays an important role in singing: some notes and note sequences are
much more probable than others when the whole performance is treated as
a musical entity. This section considers how to exploit the knowledge about
musical context in singing transcription.

Musical Key

The tonality of a singing performance is largely determined by its musical

key, which characterizes the basic note scale used within the performance
(see Chapter 1 for a discussion of tonal music). The first note of the scale is
called a tonic note, which also gives the name for the key. For example, the
C major scale consists of the notes C, D, E, F, G, A, and B (the white piano
keys), where the intervals between the notes in semitones are 2, 2, 1, 2, 2,
and 1, respectively. The natural A minor scale consists of the notes A, B, C,
D, E, F, and G, where 2, 1, 2, 2, 1, and 2 are the corresponding intervals.
Since C major and natural A minor scales actually contain the same notes,
the corresponding keys (C major and A minor) are referred to as relative keys
(or the relative-key pair). In general, the keys are relative if they correspond
to a major and a natural minor scale with the same notes.
The term pitch class stems from the octave-equivalence of notes. That is,
notes separated by an octave (or several octaves) belong to the same pitch
class, so that for example C is the pitch class of C3, C4, and C5. Krumhansl
reported the occurrence frequencies of different pitch classes with respect to
the tonic note of a piece, measured from a large amount of classical music
[377]. The distributions are shown in Fig. 12.12, where the pitch-class names
are listed for the relative keys C major and A minor. As can be observed.
386 Matti Ryynanen

>
o
MAJOR KEY MINOR KEY

g 0.2 RELATIVE MINOR 0.2 RELATIVE MAJOR

3
O KEY TONIC KEY TONIC
^0.15 0.15
LL
LU
O 0.1 0.1
z
LU
g^ 0.05
D
o 4
E
i.l.i.i
5
F
7
G A
8
B
9 10 11
0.05

D
4 5
Lu
8
F
9 10 11
G
8 DISTANCE FROM THE TONIC PITCH CLASS (SEMITONES)

Fig. 12.12. Pitch class occurrence frequencies in major and minor keys with respect
to the tonic pitch class (after Krumhansl [377, p. 67]). As an example, the pitch class
names are listed below the figure axes for the relative keys C major and A minor.

pitch classes such as C, E, and A often occur in both of these keys, and the
pitch classes which belong to the scale of the key are much more frequent
than the others. To conclude, given the key of a singing performance, we may
determine the probabilities of different pitch classes from a musical point of
view and use this knowledge to solve ambiguous note labellings.

A musical key estimation method

Viitaniemi et al. proposed a key estimation method which produces the prob-
abilities of different keys given a sequence of FO estimates [643]. The required
prior knowledge consists of the occurrence probabilities of different notes n
given a key k, P{n \k). As only the pitch class of a note is assumed to affect
the mentioned probability, distributions such as those shown in Fig. 12.12 can
be used. Then the probability of a key given a note is obtained using Bayes's
formula:
P{k I n) - P{n I k)P{k) (12.8)
P{n)
Further, the authors used singing performances as training data to estimate
the probabilities of different FO values given a note n, i.e., P{FQ |n), which
was modelled with a GMM. This was to represent singing pitch deviation from
the nominal note pitch, and the FO estimation errors. Then the probability of
a key given an FO estimate was calculated as

P(fc|Fo)
P ( F o I k)P{k) (assumpt.;
E P{Fo I n)Pin I k) m_
P{Fo)
all notes n

where it was assumed that the probability of an FO estimate is independent of

the key when the note is given, i.e., P{Fo \n,k) = P{Fo \ n). The probability
of key k given a sequence of FO estimates, O [FoiX), , F()(T)), was then
calculated as
 12 Singing Transcription 387
T
P{k\0) = \{P{k\F^{t)). (12.9)

The method was reported to find the correct relative-key pair in 86% of singing
performances [643]. The estimated key probabihties were appHed as a musi-
cological model to favour certain notes in singing transcription. At time t, the
probability of note rit given the sequence of FO estimates up to time t was
defined as
24

P{nt \0) = Y. ^(^^ I ^)^(^ I ^) ' (^2-1^)

k=l

where the index k sums over all the 24 major and minor keys. It should be
noted that the above equation gives the probability of note n at time t from
the viewpoint of the musicological model alone, without any special emphasis
to the most recent observation Fo(t).

Note AT-Grams

Note AT-grams formulate the idea that the probability of a note depends on
the previous A^ 1 notes. The probability of a note rit given the previous notes
ni:t-i is here denoted by P{nt | ni:t_i). A^-grams are based on (A^l)-th order
Markov assumption which can be written as

P{nt I ni:t_i) - P{nt \ Ut-N^i-.t-i) (12.11)

The note A^-gram probabilities are estimated from databases containing note
sequences. In practice, the databases do not include all the possible note
sequences, and therefore a method is usually employed to estimate also
the probabilities of the non-occurring note sequences. In general, this process
is called smoothing. For different smoothing methods, see [321].
Ryynanen and Klapuri applied note A^-gram probabilities for A^ G {2,3}
under a given key to control the transitions between different note HMMs
[558] (see Fig. 12.11). The probability of note rit at time t was defined as
^ ( ^ t l^prev,^:), where k denotes the key and riprev = '^t-N^i-.t-i is used to
denote the N 1 previous notes for convenience. The probabilities were es-
timated by counting the occurrences of different note sequences in a large
database of monophonic MIDI files and by smoothing them with the Witten-
Bell discounting algorithm [673].
The estimated note AT-gram probabilities were applied in singing transcrip-
tion as follows. First, the major and the minor keys /cmaj, ^min of the most
probable relative key-pair were determined from the singing performance us-
ing the key estimation method of Viitaniemi et al. [643]. Then the probability
of moving to note rit was obtained by

P{nt I nprev) = ^ ( ^ ^ I ^P^ev, femaj) + Pjut \ rip.ev, fe^in) ^^^.12)

388 Matti Ryynanen

If the key information was not available, the probability was given by

1 ^^
P{nt\nprev) = ^Yl -^Kl^prev, k) , (12.13)
k=l

that is, assuming all the major and minor keys equally probable. If the mu-
sicological model was completely disabled, equal transition probabilities were
used for all note transitions.
Simulation experiments with the described musicological model showed
that the use of note A/'-grams without the key information did not improve
transcription results compared to the situation where the musicological model
was completely disabled. When using the key information with the note N-
grams, however, the error rate was reduced from 13% to under 10%, where the
note bigrams {N = 2) worked slightly better than the note trigrams {N = 3)
[558], [557].

Metrical Context

Metrical context refers to the temporal arrangement of notes with respect to

the musical tempo of the performance. Both the note beginnings and their
durations are usually related to the tempo. As an example of the note dura-
tions, given a tempo of 100 beats per minute, the duration of quarter notes is
around 60 s / 100 ^ 0.6 s, a half of this for eighth notes, two times the duration
of quarter notes for the half notes, and so forth. The note beginnings, on the
other hand, are usually located at the positions of the beats of an underlying
pulse of the performance (metrically strong positions). For a further discus-
sion on musical metre analysis, see Chapter 4. The singing notes, however, are
not necessarily performed at a constant tempo due to the expressive nature
of singing performances.
Viitaniemi et al. tested the explicit modelling of note durations within
an HMM [643]. They weighted the transitions between HMM states (corre-
sponding to different notes) by note-duration likelihoods estimated from a
large amount of MIDI files. However, it was reported that the duration mod-
elling did not improve transcription results. Ryynanen and Klapuri estimated
the tempo during singing performances and derived a metrical accent function
based on the estimated tempo [558]. The metrical accent function was used
as an acoustic feature in the note event model in order to predict note begin-
nings in singing performances. However, the advantage of using the metrical
accent was found to be insignificant compared to the increased complexity of
the transcription system.

Summary

To conclude, the utilization of musical context can significantly improve the

accuracy of singing transcription systems, although the techniques for musical
 12 Singing Transcription 389

context modelling are still under development. If the musical key or the tempo
of a performance is known in advance, the quality of the transcription is very
likely to be improved. Use of the tonal context for pitch labelling appears to
be more important than the use of the metrical context for note segmenta-
tion. However, the metrical context becomes essential if in addition the note
durations are quantized, as described in Chapter 4.

12.4.4 Post-Processing and Encoding Expression

Note-sequence post-processing can be used to encode the expressive aspects of

a singing performance into the transcription, or to perform time quantization.
The term time quantization refers to the process of associating note onsets
and durations with a discrete temporal grid. Time quantization is needed,
for example, in singing-to-score types of applications. A reader interested in
time quantization is referred to the work of Cemgil et al. [75], [77], and to
Chapter 4 in this volume.
The focus in this section is on the expression encoding. Singing perfor-
mances are often quite expressive in nature, and note sequences are insufficient
representations of singing performances in this respect. If singing transcription
is used in an application where the transcribed note sequence is resynthesized,
the plain note sequence sounds rather dull. Therefore, the transcriptions for
this type of applications should contain additional data representing the ex-
pressive aspects of singing performances, such as glissandi, vibrato, tremolo,
and loudness changes during the notes. In addition, traditional music notation
provides some limited number of expression symbols, such as loudness indica-
tors {pp, p, mf, ff). These can be derived from the estimated parameters of
expression and added to the score.
Vibrato is a very important characteristic of singing sounds. Research on
the computational analysis of vibrato has addressed the detection, estimation,
or synthesis of vibrato in singing signals [284], [461], [337]. Herrera and Bonada
detected vibrato in the FO trajectory during a note as follows [284]. Given a
note segment, the global mean of the fundamental frequency trajectory during
the note is first removed to obtain a zero-mean FO trajectory. The trajectory
is then analysed in half-overlapping windows of length 0.37 s. For each window,
the FFT of the windowed FO trajectory is computed and prominent spectral
peaks are subjected to parabolic interpolation. If a prominent peak around
5-6 Hz is found in the spectrum, the frequency and the amplitude of the peak
are interpreted as the vibrato rate and the depth in the analysed window,
respectively. The method can be straightforwardly used to detect tremolo, too,
by replacing the fundamental frequencies with framewise energy measures.
Once the vibrato parameters have been detected during a note, they can
be encoded as pitch-bend MIDI messages in the MIDI representation of the
performance.
The loudness of a note over its duration can be directly measured with
framewise energies in the note segment. Each MIDI note-on message includes
390 Matti Ryynanen

a velocity parameter which indicates the loudness of the note. To decide the
velocity value of an entire note, the RMS energies of voiced frames can be
averaged and mapped to a velocity value, for example.
The note event models which consider different types of time segments
during the singing notes allow different measurements to be made within
the different segments of transcribed notes. The amount of glissando can
be measured during the attack segment, vibrato and tremolo during the
sustain segment, and so forth. The estimation and encoding of expres-
sion parameters improves the quality of resynthesized singing transcrip-
tions considerably. Some examples on expression encoding can be found at
http://www.cs.tut.fi/sgn/axg/matti/demos/monomel.

12.5 S u m m a r y and Discussion

In this chapter, we have considered the automatic conversion of mono-

phonic singing performances into musical transcriptions and covered the basic
methodology for solving this problem. In the current singing transcription sys-
tems, frame-level features are first extracted from a singing waveform and the
features are then segmented into note events and assigned pitch labels. In the
literature, the latter two operations have been performed either consecutively
(the segment-and-label approach) or jointly with statistical methods.
In both of these approaches, it is important to also consider the musical
context, since music transcription in general is based on the perception of
musical entities at many levels rather than on the local evidence of individual
notes only. The musicological modelling has been addressed using key esti-
mation, note A/'-grams, and metrical context where especially the use of tonal
context has brought improvements in terms of transcription accuracy.
Singing transcription can be accomphshed quite satisfactorily with the
state-of-the-art methods, and these are already applicable in real-world situ-
ations. However, the quality of singing transcription is still far from perfect
as compared to the superior quality of hand-made transcriptions. To design
better methods, we need a deeper understanding of the perception of singing
sounds, and of the influence of the musical context on this. Further develop-
ment of both the acoustic and the musicological models is likely to improve
the transcription quality. This inevitably leads to the development of complex
computational models and provides interesting challenges for the future.
References

1. S. Abdallah and M. Plumbley. Unsupervised onset detection: A probabilis-

tic approach using ICA and a hidden Markov classifier. In Cambridge Music
Processing Colloquium, Cambridge, UK, 2003.
2. S.A. Abdallah. Towards Music Perception by Redundancy Reduction and Unsu-
pervised Learning in Probabilistic Models. PhD thesis, Department of Electronic
Engineering, King's College London, 2002.
3. S.A. Abdallah and M.D. Plumbley. If the independent components of natural
images are edges, what are the independent components of natural sounds?
In International Conference on Independent Component Analysis and Signal
Separation, pp. 534-539, San Diego, USA, 2001.
4. S.A. Abdallah and M.D. Plumbley. An independent component analysis ap-
proach to automatic music transcription. In Audio Engineering Society 114th
Convention, Amsterdam, Netherlands, March 2003.
5. S.A. Abdallah and M.D. Plumbley. Polyphonic transcription by non-negative
sparse coding of power spectra. In International Conference on Music Infor-
mation Retrieval, pp. 318-325, Barcelona, Spain, October 2004.
6. M. Abe and S. Ando. Application of loudness/pitch/timbre decomposition
operators to auditory scene analysis. In IEEE International Conference on
Acoustics, Speech, and Signal Processing, pp. 2646-2649, Atlanta, USA, 1996.
7. T. Abe, T. Kobayashi, and S. Imai. Robust pitch estimation with harmonics
enhancement in noisy environments based on instantaneous frequency. In Inter-
national Conference on Spoken Language Processing, pp. 1277-1280, Philadel-
phia, USA, 1996.
8. M. Abramowitz and LA. Stegun. Handbook of Mathematical Functions. Dover
Publications, New York, USA, 1970.
9. K. Achan, S. Roweis, and B. Frey. Probabilistic inference of speech signals from
phaseless spectrograms. In Neural Information Processing Systems, pp. 1393-
1400, Vancouver, Canada, 2003.
10. G. Agostini. Sistemi informatici per il riconoscimento automatico di timbriche
musicali in segnali monofonici mediante tecniche di statistica multivariata.
Technical report, Departimento di Elettronica e Informazione, Facolta di In-
gegneria, Politecnico di Milano, 1999. Tesi di Laurea.
392 References

11. G. Agostini, M. Longari, and E. Pollastri. Musical instrument timbres classifi-

cation with spectral features. In IEEE Fourth Workshop on Multimedia Signal
Processing, pp. 97-102, Cannes, France, 2001.
12. G. Agostini, M. Longari, and E. Pollastri. Musical instrument timbres classifi-
cation with spectral features. EURASIP Journal of Applied Signal Processing,
1:5-14, 2003.
13. W.A. Ainsworth, editor. Advances in Speech, Hearing and Language Processing.
JAI Press, Greenwich, Connecticut, 1996.
14. P.E. Allen and R.B. Dannenberg. Tracking musical beats in real time. In In-
ternational Computer Music Conference, Glasgow, Scotland, 1990.
15. M. Alonso, M. Badeau, B. David, and G. Richard. Musical tempo estimation
using noise subspace projections. In IEEE Workshop on Applications of Signal
Processing to Audio and Acoustics, New Paltz, USA, October 2003.
16. M. Alonso, B. David, and G. Richard. Tempo and beat estimation of mu-
sical signals. In International Conference on Music Information Retrieval,
Barcelona, Spain, October 2004.
17. B. Alpert. A class of bases in l^ for the sparse representation of integral oper-
ators. SIAM Journal of Mathematical Analysis, 24(l):246-262, 1993.
18. B.D.O. Anderson and J.B. Moore. Optimal Filtering. Prentice Hall, Englewood
Cliffs, USA, 1979.
19. C. Andrieu, M. Davy, and A. Doucet. Efficient particle filtering for jump
Markov Systems: Application to time-varying autoregressions. IEEE Trans-
actions on Signal Processing, 51 (7): 1762-1769, July 2003.
20. C. Andrieu, P.M. Djuric, and A. Doucet. Model selection by MCMC compu-
tation. Signal Processing, 81(1): 19-37, January 2001.
21. C. Andrieu and A. Doucet. Joint Bayesian detection and estimation of noisy
sinusoids via reversible jump MCMC. IEEE Transactions on Signal Processing,
47(10):2667-2676, 1999.
22. M.S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp. A tutorial on particle
filters for online nonlinear/non-Gaussian Bayesian tracking. IEEE Transactions
on Signal Processing, 50(2): 174-88, February 2002.
23. J.-J. Aucouturier and M. Sandler. Finding repeating patterns in acoustic mu-
sical signals: Applications for audio thumbnailing. In AES International Con-
ference on Virtual, Synthetic and Entertainment Audio, pp. 412-421, Espoo,
Finland, June 2002.
24. C. Avendano. Frequency-domain source identification and manipulation in
stereo mixes for enhancement, suppression and re-panning applications. In
IEEE Workshop on Applications of Signal Processing to Audio and Acoustics,
New Paltz, USA, 2003.
25. F.R. Bach and M.I. Jordan. Discriminative training of hidden Markov models
for multiple pitch tracking. In IEEE International Conference on Audio, Speech
and Signal Processing, Philadelphia, USA, March 2005.
26. Y. Bar-Shalom. Tracking methods in a multitarget environment. IEEE Trans-
actions on Automatic Control, 23(4):618-626, August 1978.
27. M.A. Bartsch and G.H. Wakefield. To catch a chorus: Using chroma-based
representations for audio thumbnailing. In IEEE Workshop on Applications of
Signal Processing to Audio and Acoustics, pp. 15-18, New Paltz, USA, 2001.
28. A.J. Bell and T.J. Sejnowski. An information-maximization approach to blind
separation and blind deconvolution. Neural Computation, 7:1129-1159, 1995.
 References 393

29. S.D. Bella and I. Peretz. Music agnosias: Selective impairments of music recog-
nition after brain damage. Journal of New Music Research^ 28(3):209-216,
1999.
30. J.P. Bello, L. Daudet, S. Abdallah, C. Duxbury, M. Davies, and M.B. Sandler.
A tutorial on onset detection in music signals. IEEE Transactions on Speech
and Audio Processing^ 2005. (in press).
31. J.P. Bello and M. Sandler. Phase-based note onset detection for music signals.
In IEEE International Conference on Acoustics, Speech, and Signal Processing^
Hong Kong, China.
32. L. Benaroya, F. Bimbot, L. McDonagh, and R. Gribonval. Non-negative sparse
representation for Wiener based source separation with a single sensor. In IEEE
International Conference on Acoustics, Speech, and Signal Processing^ Hong
Kong, China, 2003.
33. A.L. Berenzweig and D.P.W. Ellis. Locating singing voice segments within
music signals. In IEEE Workshop on Applications of Signal Processing to Audio
and Acoustics, pp. 119-122, New Paltz, USA, October 2001.
34. J. Berger, R. Coifman, and M. Goldberg. Removing noise from music using lo-
cal trigonometric bases and wavelet packets. Journal of the Audio Engineering
Society, 42(10):808-818, 1994.
35. D.P. Bertsekas. Nonlinear Programming. Athena Scientific, Belmont, Massa-
chusetts, 1995.
36. E. Bigand. Contributions of music to research on human auditory cognition.
In S. McAdams and E. Bigand, editors. Thinking in Sound: The Cognitive
Psychology of Human Audition, pp. 231-277. Oxford University Press, 1993.
37. J.A. Bilmes. Timing is of the Essence: Perceptual and Computational Tech-
niques for Representing, Learning, and Reproducing Expressive Timing in Per-
cussive Rhythm. Master's thesis, Massachusetts Institute of Technology, Sep-
tember 1993.
38. J. A. Bilmes. A gentle tutorial of the EM algorithm and its appHcation to para-
meter estimation for Gaussian mixture and hidden Markov models. Technical
report. International Computer Science Institute, Berkeley, USA, 1998.
39. E. Bingham and H. Mannila. Random projection in dimensionality reduction:
AppKcations to image and text data. In The Seventh A CM SIGKDD Interna-
tional Conference on Knowledge Discovery and Data Mining, pp. 245-250, San
Francisco, California, 2001.
40. C M . Bishop. Neural Networks for Pattern Recognition. Oxford University
Press, Oxford, England, 1995.
41. S.S. Blackman and R. Popoli. Design and Analysis of Modem Tracking Sys-
tems, Artech House, 1999.
42. A. Blum and P. Langley. Selection of relevant features and examples in machine
learning. Artificial Intelligence, 97(l-2):245-271, 1997.
43. T. Blumensath and M. Davies. Unsupervised learning of sparse and shift-
invariant decompositions of polyphonic music. In IEEE International Con-
ference on Acoustics, Speech, and Signal Processing, Montreal, Canada, 2004.
44. J. Bonada and A. Loscos. Sample-based singing voice synthesizer by spectral
concatenation. In Stockholm Music Acoustics Conference, Stockholm, Sweden,
2003.
45. I. Borg and P. Groenen. Modem Multidimensional Scaling: Theory and Appli-
cations. Springer, 1997.
394 References

46. G.C. Bowker and S.L. Star. Sorting Things Out: Classification and Its Conse-
quences. MIT Press, Cambridge, USA, 1999.
47. G.E.P. Box and D.R. Cox. An analysis of transformations. Journal of the Royal
Statistical Society: Series B, 26:296-311, 1964.
48. R. Bracewell. Fourier Transform and Its Applications. McGraw-Hill, 1999.
49. A.S. Bregman. Auditory Scene Analysis. MIT Press, Cambridge, USA, 1990.
50. A.S. Bregman. Constraints on computational models of auditory scene analysis,
as derived from human perception. Journal of the Acoustical Society of Japan
(E), 16(3): 133-136, May 1995.
51. L. Breiman. Random forests. Machine Learning, 45:5-32, 2001.
52. G.J. Brown. Computational Auditory Scene Analysis: A Representational Ap-
proach. PhD thesis. University of Sheffield, Sheffield, UK, 1992.
53. G.J. Brown and M. Cooke. Perceptual grouping of musical sounds: A compu-
tational model. Journal of New Music Research, 23(1):107-132, 1994.
54. J.C. Brown. Musical fundamental frequency tracking using a pattern recogni-
tion method. Journal of the Acoustical Society of America, 92(3): 1394-1402,
1992.
55. J.C. Brown. Determination of the meter of musical scores by autocorrelation.
Journal of the Acoustical Society of America, 94(4): 1953-1957, October 1993.
56. J.C. Brown. Computer identification of musical instruments using pattern
recognition with cepstral coefficients as features. Journal of the Acoustical So-
ciety of America, 105:1933-1941, 1999.
57. J.C. Brown, O. Houix, and S. McAdams. Feature dependence in the automatic
identification of musical woodwind instruments. Journal of the Acoustical So-
ciety of America, 109:1064-1072, 2001.
58. J.C. Brown and P. Smaragdis. Independent component analysis for automatic
note extraction from musical trills. Journal of the Acoustical Society of Amer-
ica, 115, May 2004.
59. C.J.C. Burges. A tutorial on support vector machines for pattern recognition.
Data Mining and Knowledge Discovery, 2(2): 121-167, 1998.
60. E.M. Burns. Intervals, scales, and tuning. In Deutsch [144], pp. 215-264.
61. C.S. Burrus. Multiband least square FIR filter design. IEEE Transactions on
Signal Processing, 43(2):412-421, February 1995.
62. P. Cano, M. Koppenberger, S. Le Groux, J. Ricard, P. Herrera, and N. Wack.
Nearest-neighbor automatic sound annotation with a wordnet taxonomy. Jour-
nal of Intelligent Information Systems, 24:99-111, 2005.
63. J.-F. Cardoso. Source separation using higher order moments. In IEEE Interna-
tional Conference on Acoustics, Speech, and Signal Processing, pp. 2109-2112,
Glasgow, UK, 1989.
64. J.-F. Cardoso. Multidimensional independent component analysis. In IEEE
International Conference on Acoustics, Speech, and Signal Processing, Seattle,
USA, 1998.
65. J.-F. Cardoso. High-order contrasts for independent component analysis.
Neural Computation, 11(1), 1999.
66. P. Cariani. Recurrent timing nets for auditory scene analysis. In International
Joint Conference on Neural Networks, Portland, Oregon, July 2003.
67. P.A. Cariani and B. Delgutte. Neural correlates of the pitch of complex tones.
I. Pitch and pitch salience. II. Pitch shift, pitch ambiguity, phase invariance,
pitch circularity, rate pitch, and the dominance region for pitch. Journal of
Neurophysiology, 76(3): 1698-1734, 1996.
 References 395

68. R.P. Carlyon. Temporal pitch mechanisms in acoustic and electric hearing.
Journal of the Acoustical Society of America, 112(2):621-633, 2002.
69. F. Carreras, M. Leman, and M. Lesaffre. Automatic description of musical sig-
nals using schema-based chord decomposition. Journal of New Music Research,
28(4):310-331, 1999.
70. M.A. Casey. Auditory Group Theory with Applications to Statistical Basis
Methods for Structured Audio. PhD thesis, Massachusetts Institute of Tech-
nology, 1998.
71. M.A. Casey. General sound classification and similarity in MPEG-7. Organized
Sound, 6:153-164, 2001.
72. M.A. Casey. MPEG-7 sound-recognition tools. IEEE Transactions on Circuits
and Systems for Video Technology, 11(6), June 2001.
73. M.A. Casey and A. Westner. Separation of mixed audio sources by indepen-
dent subspace analysis. In International Computer Music Conference, Berlin,
Germany, 2000.
74. A. Cemgil. Bayesian Music Transcription. PhD thesis, Nijmegen University,
2004.
75. A.T. Cemgil, P. Desain, and B. Kappen. Rhythm quantization for transcrip-
tion. Computer Music Journal, 24(2):60-76, 2000.
76. A.T. Cemgil and B. Kappen. Tempo tracking and rhythm quantization by
sequential Monte Carlo. In Neural Information Processing Systems, Vancouver,
British Columbia, Canada, 2001.
77. A.T. Cemgil and B. Kappen. Monte Carlo methods for tempo tracking and
rhythm quantization. Journal of Artificial Intelligence Research, 18:45-81,
2003.
78. A.T. Cemgil, B. Kappen, and D. Barber. A generative model for music tran-
scription. IEEE Transactions on Speech and Audio Processing, 13(6), 2005.
79. A.T. Cemgil, B. Kappen, P. Desain, and H. Honing. On tempo tracking: Tem-
pogram representation and Kalman filtering. Journal of New Music Research,
28(4):259-273, 2001.
80. C. Chafe and D. Jaffe. Source separation and note identification in polyphonic
music. In IEEE International Conference on Acoustics, Speech, and Signal
Processing, pp. 1289-1292, Atlanta, USA, 1986.
81. C. Chafe, J. Kashima, B. Mont-Reynaud, and J. Smith. Techniques for note
identification in polyphonic music. In International Computer Music Confer-
ence, pp. 399-405, Vancouver, Canada, 1985.
82. W. Chai and B. Vercoe. Structural analysis of musical signals for indexing and
thumbnailing. In ACM/IEEE Joint Conference on Digital Libraries, pp. 27-34,
Texas, USA, 2003.
83. G. Charbonneau. Timbre and the perceptual effects of three types of data
reduction. Organized Sound, 5:10-19, 1981.
84. F.J. Charpentier. Pitch detection using the short-term phase spectrum. In
IEEE International Conference on Acoustics, Speech, and Signal Processing,
pp. 113-116, Tokyo, Japan, 1986.
85. G. Chechik, A. Globerson, M.J. Anderson, E.D. Young, I. Nelken, and
N. Tishby. Group redundancy measures reveal redundancy reduction in the au-
ditory pathway. In Neural Information Processing Systems, Vancouver, British
Columbia, Canada, 2001.
86. C.E. Cherry. Some experiments on the recognition of speech, with one and with
two ears. Journal of the Acoustical Society of America, 25(5):975-979, 1953.
396 References

87. N. Chetry, M. Davies, and M. Sandler. Identification of monophonic instru-

ment recordings using k-means and support vector machines. In Digital Music
Research Network Conference, Glasgow, England, 2005.
88. N. Chetry, M. Davies, and M Sandler. Musical instrument identification using
Isf and k-means. In Audio Engineering Society 118th Convention, Barcelona,
Spain, 2005.
89. C.K. Chui. Wavelets: A Tutorial in Theory and Applications. Academic Press,
1992.
90. K.W. Church and W.A. Gale. A comparison of the enhanced Good-Turing
and deleted estimation methods for estimating probabilities of English speech.
Computer Speech and Language, 5(1): 19-54, 1991.
91. G. Ciuperca, A. Ridolfi, and J. Idler. Penalized maximum likelihood estimator
for normal mixtures. Scandinavian Journal of Statistics, 30(l):45-59, March
2003.
92. L.P. Clarisse, J.P. Martens, M. Lesaffre, B. De Baets, H. De Meyer, and M. Le-
man. An auditory model based transcriber of singing sequences. In Interna-
tional Conference on Music Information Retrieval, Paris, Prance, October 2002.
93. E.F. Clarke. Rhythm and timing in music. In Deutsch [144], pp. 473-500.
94. D. Cliff. Hang the DJ: Automatic sequencing and seamless mixing of dance-
music tracks. Technical Report HPL-2000-104, HP Laboratories Bristol, Au-
gust 2000.
95. N. Collins. A comparison of sound onset detection algorithms with empha-
sis on psychoacoustically motivated detection functions. In Audio Engineering
Society 118th Convention, Barcelona, Spain, May 2005.
96. P. Comon. Independent component analysis - a new concept? Signal Processing,
36(3):287-314, April 1994.
97. M. Cooke. Modelling Auditory Processing and Organisation. PhD thesis. Uni-
versity of Sheffield, Sheffield, UK, 1991.
98. M. Cooke and D.P.W. Ellis. The auditory organization of speech and other
sources in listeners and computational models. Speech Communication, 35(3-
4):141-177, 2001.
99. M. Cooke, P. Green, L. Josifovski, and A. Vizinho. Robust automatic speech
recognition with missing and unreliable acoustic data. Speech Communication,
34:267-285, 2001.
100. M.P. Cooke. Modeling Auditory Processing and Organisation. Cambridge Uni-
versity Press, Cambridge, UK, 1993.
101. G. Cooper. The computational complexity of probabilistic inference using
Bayesian belief networks. Artificial Intelligence, 42(2-3):393-405, 1990.
102. G. Cooper and L.B. Meyer. The Rhythmic Structure of Music. University of
Chicago Press, 1960.
103. M. Cooper and J. Foote. Automatic music summarization via similarity analy-
sis. In International Conference on Music Information Retrieval, pp. 81-85,
Paris, France, October 2002.
104. M. Cooper and J. Foote. Summarizing popular music via structural similarity
analysis. In IEEE Workshop on Applications of Signal Processing to Audio and
Acoustics, pp. 127-130, New Paltz, USA, 2003.
105. P. Cosi, G. De Poh, and G. Lauzzana. Auditory modelhng and self-organizing
neural networks for timbre classification. Journal of New Music Research,
21:71-98, 1994.
 References 397

106. H. Cottereau, J.-M. Fiasco, C. Doncarli, and M. Davy. Two approaches for the
estimation of time-varying ampUtude multichirp signals. In IEEE International
Conference on Acoustics, Speech, and Signal Processing, Hong Kong, China,
April 2003.
107. M.S. Grouse, R.D. Nowak, and R.G. Baraniuk. Wavelet-based signal processing
using hidden Markov models. IEEE Transactions on Signal Processing, 46:886-
902, April 1998. Special issue on filter banks.
108. R. Cusack and R.P. Carlyon. Auditory perceptual organization inside and out-
side the laboratory. In J.G. Neuhoff, editor, Ecological Psychoacoustics, pp. 15-
48. Elsevier Academic Press, 2004.
109. R.B. Dannenberg, W.P. Birmingham, G. Tzanetakis, G. Meek, N. Hu, and
B. Pardo. The MUSART testbed for query-by-humming evaluation. In Inter-
national Conference on Music Information Retrieval, pp. 41-47, Baltimore,
USA, October 2003.
110. R.B. Dannenberg and N. Hu. Discovering musical structure in audio recordings.
In International Conference on Music And Artificial Intelligence, pp. 43-57,
Edinburgh, Scotland, UK, September 2002.
111. R.B. Dannenberg and N. Hu. Pattern discovery techniques for music audio.
Journal of New Music Research, 32(2):153-163, 2003.
112. T. Dasu and T. Johnson. Exploratory data mining and data cleaning. John
Wiley Sz Sons, second edition, 2003.
113. I. Daubechies. Ten Lectures on Wavelets. SIAM, Philadelphia, PA, 1992.
114. L. Daudet, S. Molla, and B. Torresani. Towards a hybrid audio coder. In Jian-
ping Li, editor. Wavelets and Its Applications, Proceedings of a Conference
Held in Chongqing (China). World Scientific Publishing Gompany, 2004.
115. L. Daudet and B. Torresani. Hybrid representations for audiophonic signal
encoding. Signal Processing, 82(11):1595-1617, 2002. Special issue on image
and video coding beyond standards.
116. Laurent Daudet. Sparse and structured decompositions of signals with the
molecular matching pursuit. IEEE Transactions on Acoustics, Speech, and Sig-
nal Processing, 2004. (in press).
117. W.B. Davenport and W.L. Root. An Introduction to the Theory of Random
Signals and Noise. IEEE Press, New York, 1987.
118. M.E. Davies and M.D. Plumbley. Gausal tempo tracking of audio. In Interna-
tional Conference on Music Information Retrieval, pp. 164-169, 2004.
119. G. Davis, S. Mallat, and M. Avellaneda. Adaptive greedy approximations. Con-
structive Approximation, 13:57-98, 1997.
120. M. Davy. Bayesian separation of harmonic sources. In Joint Statistical Meeting
of the American Statistial Association, Minneapolis, USA, August 2005.
121. M. Davy, G. Doncarli, and J.Y. Tourneret. Glassification of chirp signals using
hierarchical Bayesian learning and MGMG methods. IEEE Transactions on
Signal Processing, 50(2):377-388, 2002.
122. M. Davy and S. Godsill. Bayesian harmonic models for musical signal analy-
sis. In Seventh Valencia International meeting Bayesian statistics 7, Tenerife,
Spain, June 2002.
123. M. Davy and S. Godsill. Detection of abrupt spectral changes using support
vector machines: An application to audio signal segmentation. In IEEE Inter-
national Conference on Acoustics, Speech, and Signal Processing, Volume 2,
pp. 1313-1316, Orlando, USA, May 2002.
398 References

124. M. Davy, S. Godsill, and J. Idler. Bayesian analysis of polyphonic western tonal
music. Journal of the Acoustical Society of America^ 2005. (in press).
125. M. Davy and S.J. Godsill. Audio information retrieval: A bibliographical
study. Technical Report CUED/F-INFENG/TR.429, Department of Engineer-
ing, University of Cambridge, February 2002.
126. M. Davy and S.J. Godsill. Bayesian harmonic models for musical pitch estima-
tion and analysis. Technical Report CUED/F-INFENG/TR.431, Department
of Engineering, University of Cambridge, Cambridge, UK, April 2002.
127. M. Davy and J. Idler. Fast MCMC computations for the estimation of
sparse processes from noisy observations. In IEEE International Conference
on Acoustics, Speech, and Signal Processing^ Montreal, Canada, May 2004.
128. E. de Boer and H.R. de Jongh. On cochlear encoding: Potentials and Hmita-
tions of the reverse-correlation technique. Journal of the Acoustical Society of
America, 63(1):115-135, 1978.
129. A. de Cheveigne. Separation of concurrent harmonic sounds: Fundamental fre-
quency estimation and a time-domain cancellation model for auditory process-
ing. Journal of the Acoustical Society of America, 93(6):3271-3290, 1993.
130. A. de Cheveigne. Concurrent vowel identification. III. A neural model of har-
monic interference cancellation. Journal of the Acoustical Society of America,
101(5):2857-2865, 1997.
131. A. de Cheveigne. Cancellation model of pitch perception. Journal of the
Acoustical Society of America, 103(3):1261-1271, 1998.
132. A. de Cheveigne. Pitch perception models. In Plack and Oxenham [522].
133. A. de Cheveigne and H. Kawahara. Multiple period estimation and pitch per-
ception model. Speech Communication, 27:175-185, 1999.
134. A. de Cheveigne and H. Kawahara. Comparative evaluation of FO estimation
algorithms. In 7th European Conf. Speech Communication and Technology, Aal-
borg, Denmark, 2001.
135. A. de Cheveigne and H. Kawahara. YIN, a fundamental frequency estimator for
speech and music. Journal of the Acoustical Society of America, 111(4):1917-
1930, 2002.
136. T. De Mulder, J.P. Martens, M. Lesaffre, M. Leman, B. De Baets, and
H. De Meyer. An auditory model based transcriber of vocal queries. In Inter-
national Conference on Music Information Retrieval, Baltimore, USA, 2003.
137. T. De Mulder, J.P. Martens, M. Lesaffre, M. Leman, B. De Baets, and
H. De Meyer. Recent improvements of an auditory model based front-end
for the transcription of vocal queries. In IEEE International Conference on
Acoustics, Speech, and Signal Processing, Volume 4, pp. 257-260, Montreal,
Canada, 2004.
138. A.P. Dempster, N.M. Laird, and D.B. Rubin. Maximum likelihood from in-
complete data via the EM algorithm. Journal of the Royal Statistical Society:
Series B, 39(l):l-38, 1977.
139. P. Depalle, G. Garcia, and X. Rodet. Tracking of partials for additive sound
synthesis using hidden Markov model. In IEEE International Conference on
Acoustics, Speech, and Signal Processing, Volume 1, pp. 225-228, Minneapolis,
USA, 1993.
140. P. Desain and H. Honing. The quantization of musical time: A connectionist
approach. Computer Music Journal, 13(3):56-66, 1989.
141. P. Desain and H. Honing. Computational models of beat induction: The rule-
based approach. Journal of New Music Research, 28(l):29-42, 1999.
 References 399

142. P. Desain and H. Honing. The formation of rhythmic categories and metric
priming. Perception, 32(3):241-365, 2003.
143. F. Desobry, M. Davy, and C. DoncarH. An onhne kernel change detection al-
gorithm. IEEE Transactions on Signal Processing, 2005. (in press).
144. D. Deutsch, editor. The Psychology of Music. Academic Press, San Diego,
California, 2nd edition, 1999.
145. T.G. Dietterich. Approximate statistical tests for comparing supervised classi-
fication learning algorithms. Neural Computation, 10:1895-1923, 1998.
146. T.G. Dietterich. Machine learning for sequential data: A review. In T. Caelli,
A. Amin, R.P.W. Duin, M. Kamel, and de Kidder D., editors. Structural, Syn-
tactic, and Statistical Pattern Recognition, Volume 2396 of Lecture Notes in
Computer Science, pp. 15-30. Springer-Verlag, 2002.
147. C. Dittmar and C. Uhle. Further steps towards drum transcription of poly-
phonic music. In Audio Engineering Society 116th Convention, Berlin, Ger-
many, May 2004.
148. S. Dixon. Automatic extraction of tempo and beat from expressive performa-
ces. Journal of New Music Research, 30(l):39-58, 2001.
149. S. Dixon, F. Gouyon, and G. Widmer. Towards characterisation of music
via rhythmic patterns. In International Conference on Music Information Re-
trieval, pp. 509-516, Barcelona, Spain, 2004.
150. M. Dorfler. Gabor Analysis for a Class of Signals called Music. PhD thesis,
NuHAG, University of Vienna, 2002.
151. D. Dorran and R. Lawlor. An efficient audio time-scale modification algorithm
for use in a sub-band implementation. In International Conference on Digital
Audio Effects, pp. 339-343, 2003.
152. A. Doucet, N. de Freitas, and N. Gordon. Sequential Monte Carlo Methods in
Practice. Springer, New York, USA, 2001.
153. B. Doval and X. Rodet. Estimation of fundamental frequency of musical sound
signals. In IEEE International Conference on Acoustics, Speech, and Signal
Processing, pp. 3657-3660, Toronto, Canada, 1991.
154. J.S. Downie, J. Futrelle, and D. Tcheng. The international music information
systems evaluation laboratory: Governance, access and security. In Interna-
tional Conference on Music Information Retrieval, Barcelona, Spain, 2004.
155. C. Drake, A. Penel, and E. Bigand. Tapping in time with mechanical and
expressively performed music. Music Perception, 18(l):l-23, 2000.
156. J. Drish. Obtaining calibrated probability estimates from support vector ma-
chines. Technical report, University of California, San Diego, California, USA,
June 2001.
157. S. Dubnov. Extracting sound objects by independent subspace analysis. In
22nd International Audio Engineering Society Conference, Espoo, Finland,
June 2002.
158. C. Dubois and M. Davy. Harmonic tracking in spectrograms. IEEE Transac-
tions on Signal Processing, 2005. Submitted.
159. C. Dubois and M. Davy. Harmonic tracking using sequential Monte Carlo. In
IEEE Statistical Signal Processing Workshop, Bordeaux, France, July 2005.
160. C. Dubois and M. Davy. Suivi de trajectoires temps-frequence par filtrage par-
ticulaire. In 20st GRETSI colloquium, Louvain-La-Neuve, Belgium, September
2005. In French.
161. R.O. Duda, P.E. Hart, and D.G. Stork. Pattern Classification. Wiley, New
York, USA, second edition, 2001.
400 References

162. C. Duxbury, J.P. Bello, M. Davies, and M. Sandler. A combined phase and
amplitude based approach to onset detection for audio segmentation. In 4th
European Workshop on Image Analysis for Multimedia Interactive Services,
London, UK, 2003.
163. C. Duxbury, J.P. Bello, M. Davies, and M. Sandler. Complex domain onset
detection for musical signals. In International Conference on Digital Audio
Effects, London, UK, 2003.
164. C. Duxbury, M. Sandler, and M.E. Davies. A hybrid approach to musical note
onset detection. In International Conference on Digital Audio Effects, Ham-
burg, Germany, September 2002.
165. D. Eck. A positive-evidence model for classifying rhythmical patterns. Techni-
cal Report IDSIA-09-00, Dalle Molle Institute for Artificial Intelligence, 2000.
166. D. Eck. A network of relaxation oscillators that finds downbeats in rhythms.
Technical Report IDSIA-06-01, Dalle Molle Institute for Artificial Intelligence,
2001.
167. J. Eggink and G.J. Brown. Application of missing feature theory to the recog-
nition of musical instruments in polyphonic audio. In International Conference
on Music Information Retrieval, pp. 125-131, Baltimore, USA, 2003.
168. J. Eggink and G.J. Brown. Application of missing feature theory to the recog-
nition of musical instruments in polyphonic audio. In IEEE International Con-
ference on Acoustics, Speech, and Signal Processing, pp. 553-556, Hong Kong,
China, 2003.
169. J. Eggink and G.J. Brown. Extracting melody lines from complex audio. In In-
ternational Conference on Music Information Retrieval, pp. 84-91, Barcelona,
Spain, October 2004.
170. J. Eggink and G.J. Brown. Instrument recognition in accompanied sonatas and
concertos. In IEEE International Conference on Acoustics, Speech, and Signal
Processing, pp. 217-220, Montreal, Canada, 2004.
171. D.P.W. Ellis. Prediction-Driven Computational Auditory Scene Analysis. PhD
thesis, Massachusetts Institute of Technology, 1996.
172. D.P.W. Elhs. Using knowledge to organize sound: The prediction-driven
approach to computational auditory scene analysis, and its application to
speech/nonspeech mixtures. Speech Communication, 27:281-298, 1999.
173. D.P.W. Elhs and D.F. Rosenthal. Mid-level representations for computational
auditory scene analysis. In International Joint Conference on Artificial Intel-
ligence, Montreal, Quebec, 1995.
174. A. Eronen. Automatic musical instrument recognition. Master's thesis, Tam-
pere University of Technology, 2001.
175. A. Eronen. Musical instrument recognition using ICA-based transform of fea-
tures and discriminatively trained HMMs. In Seventh International Symposium
on Signal Processing and its Applications, pp. 133-136, Paris, Prance, 2003.
176. A. Eronen and A. Klapuri. Musical instrument recognition using cepstral coef-
ficients and temporal features. In IEEE International Conference on Acoustics,
Speech, and Signal Processing, pp. 753-756, Istanbul, Turkey, 2000.
177. S. Essid, G. Richard, and B. David. Musical instrument recognition based on
class pairwise feature selection. In International Conference on Music Infor-
mation Retrieval, Barcelona, Spain, 2004.
178. S. Essid, G. Richard, and B. David. Musical instrument recognition on solo
performance. In European Signal Processing Conference, Vienna, Austria, 2004.
 References 401

179. S. Essid, G. Richard, and B. David. Inferring efficient hierarchical taxonomies

for mir tasks: AppUcation to musical instruments. In 6th International Con-
ference on Music Information Retrieval, London, UK, 2005.
180. S. Essid, G. Richard, and B. David. Instrument recognition in polyphonic mu-
sic. In IEEE International Conference on Audio, Speech and Signal Processing,
Philadelphia, USA, 2005.
181. A.M. Fanelli, G. Castellano, and C.A. Buscicchio. A modular neuro-fuzzy net-
work for musical instruments classification. In J. Kittler and F. Roli, editors,
Proceedings of the First International Workshop on Multiple Classifier Systems,
Volume 1857 of Lecture Notes In Computer Science, pp. 372-382. Springer,
2000.
182. FastICA package for MATLAB. http://www.cis.hut.fi/projects/ica/fastica/.
183. U. Fayyad and K.B. Irani. Multi-interval discretisation of continuous-valued
attributes for classification learning. In 13th International Joint Conference on
Artificial Intelligence, pp. 1022-1027, Berlin, Heidelberg, 1993.
184. H.G. Feichtinger and T. Strohmer. Gabor Analysis and Algorithms: Theory and
Applications. Birkhauser, 1998.
185. H.G. Feichtinger and T. Strohmer, editors. Advances in Gabor analysis.
Birkhauser, Boston, 2003.
186. D. FitzGerald. Automatic Drum Transcription and Source Separation. PhD
thesis, Dublin Institute of Technology, 2004.
187. D. FitzGerald, E. Coyle, and B. Lawlor. Sub-band independent subspace analy-
sis for drum transcription. In International Conference on Digital Audio Ef-
fects, Hamburg, Germany, 2002.
188. D. FitzGerald, E. Coyle, and B. Lawlor. Prior subspace analysis for drum
transcription. In Audio Engineering Society 114th Convention, Amsterdam,
Netherlands, March 2003.
189. D. FitzGerald, R. Lawlor, and E. Coyle. Drum transcription in the presence of
pitched instruments using prior subspace analysis. In Irish Signals & Systems
Conference 2003, Limerick, Ireland, July 2003.
190. D. FitzGerald, R. Lawlor, and E. Coyle. Drum transcription using automatic
grouping of events and prior subspace analysis. In 4^h European Workshop on
Image Analysis for Multimedia Interactive Services, pp. 306-309, 2003.
191. P. Flandrin. Time-Frequency/Time-Scale Analysis. Academic Press, 1999.
192. H. Fletcher. Auditory patterns. Reviews of Modem Physics, 12:47-65, 1940.
193. N.H. Fletcher and T.D. Rossing. The Physics of Musical Instruments. Springer,
Berlin, Germany, second edition, 1998.
194. J. Foote and S. Uchihashi. The beat spectrum: A new approach to rhythm
analysis. In IEEE International Conference on Multimedia and Expo, 2001.
195. J.T. Foote and M.L. Cooper. Media segmentation using self-similarity de-
composition. In SPIE Storage and Retrieval for Media Databases 2003, Vol-
ume 5021, pp. 167-175, 2003.
196. Y. Freund and R.E. Schapire. Experiments with a new boosting algorithm. In
International Conference on Machine Learning, Bari, Italy, 1996.
197. J.H. Friedman and J.W. Tukey. A projection pursuit algorithm for exploratory
data analysis. IEEE Transactions on Computers, 23:881-890, 1974.
198. Lawrence Fritts. University of Iowa Musical Instrument Samples.
http://theremin.music.uiowa.edu/MIS.html.
199. I. Fujinaga and K. MacMillan. Realtime recognition of orchestral instruments.
In International Computer Music Conference, Berlin, Germany, 2000.
402 References

200. I. Pujinaga, S. Moore, and D.S. Sullivan. Implementation of exemplar-based

learning model for music cognition. In 5th Int. Conf. on Music Perception and
Cognition, Seoul, Korea, 1998.
201. Takuya Fujishima. Realtime chord recognition of musical sound: A system us-
ing common lisp music. In International Computer Music Conference, pp. 464-
467, Beijing, China, 1999.
202. K. Fukunaga. Introduction to Statistical Pattern Recognition. Academic Press,
second edition, 1990.
203. S. Furui and M.M. Sondhi, editors. Advances in Speech Signal Processing. Mar-
cel Dekker, New York, USA, 1991.
204. T. Gautama and M.M. Van HuUe. Separation of acoustic signals using self-
organizing neural networks. In IEEE Neural Network for Signal Processing
Workshop, Madison, USA, 1999.
205. Z. Ghahramani. An introduction to hidden Markov models and Bayesian net-
works. International Journal of Pattern Recognition and Artificial Intelligence,
15(l):9-42, 2001.
206. M.A. Ghezzo. Music Theory, Ear Training, Rhythm, Solfege, and Dictation: A
Comprehensive Course. The University of Alabama Press, Alabama, 1980.
207. A. Ghias, J. Logan, D. Chamberhn, and B.C. Smith. Query by humming: Musi-
cal information retrieval in an audio database. In ACM Multimedia Conference,
pp. 231-236, San Fransisco, CaHfornia, November 1995.
208. W.R. Gilks, S. Richardson, and D.J. Spiegelhalter, editors. Markov Chain
Monte Carlo in Practice. Chapman & Hall, London, UK, 1996.
209. O. Gillet and G. Richard. Automatic transcription of drum loops. In IEEE In-
ternational Conference on Acoustics, Speech, and Signal Processing, Montreal,
Canada, May 2004.
210. O. Gillet and G. Richard. Drum loops retrieval from spoken queries. Journal
of Intelligent Information Systems, 2005.
211. O.K. Gillet and G. Richard. Automatic labeling of tabla signals. In Interna-
tional Conference on Music Information Retrieval, Baltimore, USA, October
2003.
212. S. Godsill and M. Davy. Bayesian harmonic models for musical pitch estima-
tion and analysis. In IEEE International Conference on Acoustics, Speech, and
Signal Processing, Orlando, USA, 2002.
213. S.J. Godsill and M. Davy. Bayesian computational models for inharmonicity in
musical instruments. In IEEE Workshop on Applications of Signal Processing
to Audio and Acoustics, New Paltz, USA, October 2005.
214. S.J. Godsill and P.J.W. Rayner. Digital Audio Restoration: A Statistical Model-
Based Approach. Springer, Berlin, Germany, September 1998.
215. D. Godsmark and G.J. Brown. A blackboard architecture for computational
auditory scene analysis. Speech Communication, 27(3):351-366, 1999.
216. B. Gold and N. Morgan. Speech and Audio Signal Processing: Processing and
Perception of Speech and Music. John Wiley & Sons, 2000.
217. D.E. Goldberg. Genetic Algorithms in Search, Optimization and Machine
Learning. Addison Wesley, 1989.
218. I.J. Good. The population frequencies of species and the estimation of popu-
lation parameters. Biometrika, 16(40): 16-264, 1953.
219. M. Goodwin. Residual modeling in music analysis-synthesis. In IEEE Interna-
tional Conference on Acoustics, Speech, and Signal Processing, pp. 1005-1008,
Atlanta, USA, 1996.
 References 403

220. M. Goto. A Study of Real-Time Beat Tracking for Musical Audio Signals. PhD
thesis, Waseda University, 1998.
221. M. Goto. An audio-based real-time beat tracking system for music with or
without drum-sounds. Journal of New Music Research, 30(2):159-171, 2001.
222. M. Goto. A predominant-FO estimation method for CD recordings: MAP esti-
mation using EM algorithm for adaptive tone models. In IEEE International
Conference on Acoustics, Speech, and Signal Processing, Volume 5, pp. 3365-
3368, Salt Lake City, USA, May 2001.
223. M. Goto. A predominant-FO estimation method for real-world musical audio
signals: Map estimation for incorporating prior knowledge about FOs and tone
models. In Proc. Workshop on Consistent and Reliable Acoustic Cues for Sound
Analysis, Aalborg, Denmark, 2001.
224. M. Goto. A chorus-section detecting method for musical audio signals. In
IEEE International Conference on Acoustics, Speech, and Signal Processing,
Volume 5, pp. 437-440, Hong Kong, China, April 2003.
225. M. Goto. Music scene description project: Toward audio-based real-time music
understanding. In International Conference on Music Information Retrieval,
pp. 231-232, Baltimore, USA, October 2003.
226. M. Goto. SmartMusicKIOSK: Music listening station with chorus-search func-
tion. In ACM Symposium on User Interface Software and Technology, pp. 3 1 -
40, Vancouver, British Columbia, Canada, 2003.
227. M. Goto. Development of the RWC music database. In the 18th International
Congress on Acoustics, Volume 1, pp. 553-556, Kyoto, Japan, 2004.
228. M. Goto. A real-time music scene description system: Predominant-FO esti-
mation for detecting melody and bass lines in real-world audio signals. Speech
Communication, 43(4) :311-329, 2004.
229. M. Goto, H. Hashiguchi, T. Nishimura, and R. Oka. RWC music database:
Popular, classical, and jazz music databases. In International Conference on
Music Information Retrieval, pp. 287-288, Paris, Prance, October 2002.
230. M. Goto, H. Hashiguchi, T. Nishimura, and R. Oka. RWC music database:
Music genre database and musical instrument sound database. In International
Conference on Music Information Retrieval, pp. 229-230, Baltimore, USA,
October 2003.
231. M. Goto and S. Hayamizu. A real-time music scene description system: Detect-
ing melody and bass lines in audio signals. In International Joint Conference
on Artificial Intelligence, pp. 31-40, Stockholm, Sweden, 1999.
232. M. Goto and S. Hayamizu. A real-time music scene description system: Detect-
ing melody and bass lines in audio signals. In Working Notes of the IJCAI-99
Workshop on Computational Auditory Scene Analysis, pp. 31-40, Stockholm,
Sweden, 1999.
233. M. Goto and K. Hirata. Invited review "Recent studies on music information
processing". Acoustical Science and Technology (edited by the Acoustical So-
ciety of Japan), 25(6):419-425, November 2004.
234. M. Goto and H. Muraoka. Issues in evaluating beat tracking systems. In Work-
ing Notes of the IJCAI-97 Workshop on Issues in AI and Music, pp. 9-17,
August 1997.
235. M. Goto and Y. Muraoka. A beat tracking system for acoustic signals of music.
In ACM International Conference on Multimedia, pp. 365-372, San Pransisco,
California, October 1994.
404 References

236. M. Goto and Y. Muraoka. A sound source separation system for percussion in-
struments. Transactions of the Institute of Electronics, Information and Com-
munication Engineers D-II, J77-D-II(5):901-911, May 1994. (in Japanese)
237. M. Goto and Y. Muraoka. Music understanding at the beat level: Real-time
beat tracking for audio signals. In International Joint Conference on Artificial
Intelligence, pp. 68-75, Montreal, Quebec, 1995.
238. M. Goto and Y. Muraoka. A real-time beat tracking system for audio signals.
In International Computer Music Conference, Tokyo, Japan, 1995.
239. M. Goto and Y. Muraoka. Real-time rhythm tracking for drumless audio sig-
nals: Chord change detection for musical decisions. In Working Notes of the
IJCAI-97 Workshop on Computational Auditory Scene Analysis, pp. 135-144,
Nagoya, Japan, 1997.
240. M. Goto and Y. Muraoka. Real-time beat tracking for drumless audio signals:
Chord change detection for musical decisions. Speech Communication, 27(3-
4):311-335, 1999.
241. M. Goto, M. Tabuchi, and Y. Muraoka. An automatic transcription system
for percussion instruments. In the 4^th National Convention of Information
Processing Society of Japan, 7Q-2, Tokyo, Japan, March 1993. (in Japanese)
242. F. Gouyon. Towards Automatic Rhythm Description of Musical Audio Sig-
nals: Representations, Computational Models and Applications. Master's the-
sis, UPF, Barcelona, 2003.
243. F. Gouyon and S. Dixon. A review of automatic rhythm description systems.
Computer Music Journal, 29(1), 2005.
244. F. Gouyon, L. Fabig, and J. Bonada. Rhythmic expressiveness transformations
of audio recordings: swing modifications. In International Conference on Digital
Audio Effects, 2003.
245. F. Gouyon and P. Herrera. Exploration of techniques for automatic labeling of
audio drum tracks' instruments. In MOSART Workshop on Current Research
Directions in Computer Music, Barcelona, Spain, 2001.
246. F. Gouyon and P. Herrera. Determination of the meter of musical audio signals:
Seeking recurrences in beat segment descriptors. In Audio Engineering Society
114th Convention, Amsterdam, Netherlands, March 2003.
247. F. Gouyon, P. Herrera, and P. Cano. Pulse-dependent analysis of percussive
music. In 22nd International Audio Engineering Society Conference, Espoo,
Finland, 2002.
248. F. Gouyon, A. Klapuri, S. Dixon, M. Alonso, G. Tzanetakis, C. Uhle, and
P. Cano. An experimental comparison of audio tempo induction algorithms.
IEEE Transactions on Speech and Audio Processing, 2005. (in press).
249. F. Gouyon and B. Meudic. Towards rhythmic content processing of musical
signals: Fostering complementary approaches. Journal of New Music Research,
32(1), 2003.
250. F. Gouyon, F. Pachet, and O. Delerue. On the use of zero-crossing rate for an
application of classification of percussive sounds. In International Conference
on Digital Audio Effects, Verona, Italy, December 2000.
251. P.J. Green. On use of the EM algorithm for penalized likehhood estimation.
Journal of the Royal Statistical Society: Series B, 52:443-452, 1990.
252. P.J. Green. Penalized likehhood. In Encyclopaedia of Statistical Sciences, Vol-
ume 3, pp. 578-586, 1999.
253. J.M. Grey. Multidimensional perceptual scaling of musical timbres. Journal of
the Acoustical Society of America, 61:1270-1277, 1977.
 References 405

254. J.M. Grey and J.A. Moorer. Perceptual evaluations of synthesized musical
instrument tones. Journal of the Acoustical Society of America, 62:454-462,
1977.
255. R. Gribonval. Approximations Non-lineaires pour VAnalyse des Signaux
Sonores. PhD thesis, Universite de Paris IX Dauphine, 1999.
256. R. Gribonval and E. Bacry. Harmonic decomposition of audio signals with
matching pursuit. IEEE Transactions on Signal Processing, 51(1):101-111,
2003.
257. R. Gribonval, E. Bacry, S. Mallat, Ph. Depalle, and X. Rodet. Analysis of
sound signals with high resolution matching pursuit. In IEEE Symposium on
Time-Frequency and Time-Scale Analysis, pp. 125-128, Paris, France, June
1996.
258. R. Gribonval, L. Benaroya, E. Vincent, and C. Fevotte. Proposals for perfor-
mance measurement in source separation. In the 4th International Symposium
on Independent Component Analysis and Blind Signal Separation, Nara, Japan,
2003.
259. D. Griffin and J. Lim. Signal estimation from modified short-time fourier trans-
form. IEEE Transactions on Acoustics, Speech, and Signal Processing, 32:236-
242, 1984.
260. T.D. Griffiths, J.D. Warren, S.K. Scott, I. Nelken, and A.J. King. Cortical
processing of complex sound: A way forward? Trends in Neuroscience, 27:181-
185, 1977.
261. K. Grochenig. Foundations of Time-frequency Analysis. Birkhauser, Boston,
MA, 2001.
262. I. Guyon and A. Elisseeff. An introduction to variable and feature selection.
Journal of Machine Learning Research, 3:1157-1182, 2003.
263. S.W. Hainsworth. Techniques for the Automated Analysis of Musical Audio.
PhD thesis. Department of Engineering, University of Cambridge, 2004.
264. S.W. Hainsworth and M.D. Macleod. Automatic bass line transcription from
polyphonic music. In International Computer Music Conference, pp. 431-434,
Havana, Cuba, 2001.
265. S.W. Hainsworth and M.D. Macleod. Onset detection in musical audio signals.
In International Computer Music Conference, Singapore, 2003.
266. S.W. Hainsworth and M.D. Macleod. Particle filtering applied to musical tempo
tracking. Journal of Applied Signal Processing, 15:2385-2395, 2004.
267. M.A. Hall. Correlation-Based Feature Selection for Machine Learning. PhD
thesis. Department of Computer Science, University of Waikato, Hamilton,
New Zealand, 1998.
268. M.A. Hall. Correlation-based feature selection for discrete and numeric class
machine learning. In Seventeeth International Conference on Machine Learn-
ing, Stanford, CA, USA, 2000.
269. K.N. Hamdy, A. AH, and A.H. Tewfik. Low bit rate high quality audio coding
with combined harmonic and wavelet representations. In IEEE International
Conference on Acoustics, Speech, and Signal Processing, Volume 2, pp. 1045-
1048, Atlanta, USA, 1996.
270. S. Handel. Listening: An Introduction to the Perception of Auditory Events.
MIT Press, 1989.
271. S. Handel. Timbre perception and auditory object identification. In Moore
[474], pp. 425-460.
406 References

272. A. Harma, M. Karjalainen, L. Savioja, V. Valimaki, U. K. Laine, and

J. Huopaniemi. Prequency-warped signal processing for audio applications.
Journal of the Audio Engineering Society, 48(11): 1011-1031, 2000. URL:
www.acoustics.hut.fi/software/warp/.
273. A. Harma and K. Palomaki. HUTearA free Matlab toolbox for modeling of
auditory system. In Matlab DSP conference, pp. 96-99, Espoo, Finland, 1999.
URL: www.acoustics.hut.fi/software/HUTear/.
274. F.J. Harris. On the use of windows for harmonic analysis with the discrete
Fourier transform. Proceedings of the IEEE, 66(l):51-83, January 1978.
275. W.M. Hartmann. Pitch, periodicity, and auditory organization. Journal of the
Acoustical Society of America, 100(6):3491-3502, 1996.
276. W.M. Hartmann. Signals, Sound, and Sensation. Springer, New York, 1998.
277. W.M. Hartmann, S. McAdams, and B.K. Smith. Hearing a mistuned harmonic
in an otherwise periodic complex tone. Journal of the Acoustical Society of
America, 88(4):1712-1724, 1990.
278. G. Haus and E. Pollastri. An audio front end for query-by-humming systems. In
2nd Annual International Symposium on Music Information Retrieval, Bloom-
ington, Indiana, USA, 2001.
279. S. Haykin. Neural Networks: A Comprehensive Foundation. Prentice Hall, New
Jersey, 1999.
280. D.P. Hedges. Taking Notes: The History, Practice, and Innovation of Musical
Dictation in English and American Aural Skills Pedagogy. PhD thesis. School
of Music, Indiana University, 1999.
281. T. Heittola. Automatic Classification of Music Signals. Master's thesis, Tam-
pere University of Technology, December 2003.
282. M. Helen and T. Virtanen. Separation of drums from polyphonic music us-
ing non-negtive matrix factorization and support vector machine. In European
Signal Processing Conference, Antalya, Turkey, 2005.
283. P. Herrera, X. Amatriain, E. Bat lie, and X. Serra. Towards instrument segmen-
tation for music content description: A critical review of instrument classifica-
tion techniques. In International Conference on Music Information Retrieval,
Plymouth, Massachusetts, USA, 2000.
284. P. Herrera and J. Bonada. Vibrato extraction and parametrization in the spec-
tral modehng synthesis framework. In First C0ST-G6 Workshop on Digital
Audio Effects, Barcelona, Spain, November 1998.
285. P. Herrera, A. Dehamel, and F. Gouyon. Automatic labeling of unpitched per-
cussion sounds. In Audio Engineering Society 114th Convention, Amsterdam,
Netherlands, March 2003.
286. P. Herrera, G. Peeters, and S. Dubnov. Automatic classification of musical
instrument sounds. Journal of New Music Research, 32:3-21, 2003.
287. P. Herrera, A. Yeterian, and F. Gouyon. Automatic classification of drum
sounds: A comparison of feature selection methods and classification tech-
niques. In International Conference on Music and Artificial Intelligence,
pp. 69-80, Edinburgh, Scotland, UK, September 2002.
288. P. Herrera, A. Yeterian, and F. Gouyon. Automatic classification of drum
sounds: a comparison of feature selection methods and classification techniques.
In C. Anagnostopoulou, M. Ferrand, and A. Smaill, editors, Proceedings of
the Second International Conference on Music and Artificial Intelligence, Vol-
ume 2445 of Lecture Notes In Computer Science, pp. 69-80. Springer, 2002.
 References 407

289. W.J. Hess. Pitch Determination of Speech Signals. Springer, Berlin Heidelberg,
1983.
290. W.J. Hess. Pitch and voicing determination. In Furui and Sondhi [203], pp. 3 -
48.
291. M.J. Hewitt and R. Meddis. An evaluation of eight computer models of mam-
malian inner hair-cell function. Journal of the Acoustical Society of America,
90(2):904-917, 1991.
292. T.W. Hilands and S.C.A. Thomopoulos. Nonlinear filtering methods for har-
monic retrieval and model order selection in Gaussian and non-Gaussian noise.
IEEE Transactions on Signal Processing, 45(4):982-995, April 1997.
293. J. Holland. Practical Percussion: A Guide to the Instruments and Their
Sources. Oxford University Press, 2001.
294. H. Honing. From time to time: The representation of timing and tempo. Com-
puter Music Journal, 25(3):50-61, 2001.
295. R.A. Horn and C.R. Johnson. Topics in Matrix Analysis. Cambridge University
Press, Cambridge, UK, 1994.
296. E. V. Hornbostel and C. Sachs. The classification of musical instruments. Galpin
Society Journal, pp. 3-29, 1961.
297. A.J.M. Houtsma. Pitch perception. In Moore [474], pp. 267-295.
298. A.J.M. Houtsma and J.L. Goldstein. The central origin of the pitch of complex
tones: Evidence from musical interval recognition. Journal of the Acoustical
Society of America, 51(2):520-529, 1972.
299. P. Hoyer. Non-negative sparse coding. In IEEE Workshop on Networks for
Signal Processing XII, Martigny, Switzerland, 2002.
300. P.O. Hoyer. Non-negative matrix factorization with sparseness constraints.
Journal of Machine Learning Research, 5:1457-1469, 2004.
301. N. Hu and R.B. Dannenberg. A comparison of melodic database retrieval
techniques using sung queries. In ACM/IEEE Joint Conference on Digital Li-
braries, pp. 301-307, Oregon, USA, 2002.
302. A. Hyvarinen. Fast and robust fixed-point algorithms for independent compo-
nent analysis. IEEE Transactions on Neural Networks, 10(3):626-634, 1999.
303. A. Hyvarinen and P. Hoyer. Emergence of phase and shift invariant features
by decomposition of natural images into independent feature subspaces. Neural
Computation, 12(7): 1705-1720, 2000.
304. A. Hyvarinen, J. Karhunen, and E. Oja. Independent Component Analysis.
John Wiley & Sons, 2001.
305. Vamtech Enterprises Inc. Drumtrax 3.0. Buffalo, New York, USA, 1999.
306. H. Indefrey, W. Hess, and G. Seeser. Design and evaluation of double-transform
pitch determination algorithms with nonlinear distortion in the frequency
domainpreliminary results. In IEEE International Conference on Acoustics,
Speech, and Signal Processing, pp. 415-418, Tampa, Florida, 1985.
307. International Organization for Standardization. ISO/IEC 15938-4:2002 In-
formation Technology - Multimedia Content Description Interface - Part 4-
Audio. International Organization for Standardization, Geneva, Switzerland,
2002.
308. R. Irizarry. Local harmonic estimation in musical sound signals. Journal of the
American Statistical Association, 96(454) :357-367, June 2001.
309. R. Irizarry. Weighted estimation of harmonic components in a musical sound
signal. Journal of Time Series Analysis, 23(l):29-48, 2002.
408 References

310. R.A. Irizarry. Statistics and Music: Fitting a Local Harmonic Model to Musical
Sound Signals. PhD thesis, University of Cahfornia, Berkeley, 1998.
311. F. Jaillet. Representations et Traitement Temps-frequence des Signaux Au-
dionumeriques pour des Applications de Design Sonore. PhD thesis, LATP
and Universite de Provence, Marseille, 2005.
312. F. Jaillet and B. Torresani. Time-frequency jigsaw puzzle: Adaptive multiwin-
dow and multilayered gabor expansions. Technical report, LATP and Univer-
site de Provence, Marseille, 2004. (submitted)
313. A.K. Jain, R.P.W. Duin, and J. Mao. Statistical pattern recognition: A review.
IEEE Transactions on Pattern Analysis and Machine Intelligence^ 22:4-37,
2000.
314. G.-J. Jang and T.-W. Lee. A maximum likelihood approach to single channel
source separation. Journal of Machine Learning Research, 23:1365-1392, 2003.
315. H. Jarvelainen, V. Valimaki, and M. Karjalainen. Audibility of the timbral ef-
fects of inharmonicity in stringed instrument tones. Acoustics Research Letters
Online, 2(3):79-84, 2001.
316. F. Jelinek. Statistical Methods for Speech Recognition. MIT Press, Cambridge,
Massachusetts, 1997.
317. F. Jelinek and R.L. Mercer. Interpolated estimation of Markov source para-
meters from sparse data. In Intemation Workshop on Pattern Recognition in
Practice, Amsterdam, The Netherlands, May 1980.
318. K. Jensen and T.H. Andersen. Beat estimation on the beat. In IEEE Workshop
on Applications of Signal Processing to Audio and Acoustics, New Paltz, USA,
2003.
319. K. Jensen and J. Arnspang. Binary decission tree classification of musical
sounds. In International Computer Music Conference, Beijing, China, 1999.
320. I.T. Jolliffe. Principal Component Analysis. Springer-Verlag, New York, USA,
1986.
321. D. Jurafsky and J.H. Martin. Speech and Language Processing. Prentice Hall,
New Jersey, USA, 2000.
322. C. Kaernbach and L. Demany. Psychophysical evidence against the autocorre-
lation theory of auditory temporal processing. Journal of the Acoustical Society
of America, 104(4):2298-2306, 1998.
323. T. Kageyama, K. Mochizuki, and Y. Takashima. Melody retrieval with hum-
ming. In International Computer Music Conference, pp. 349-351, Tokyo,
Japan, 1993.
324. H. Kameoka, T. Nishimoto, and S. Sagayama. Separation of harmonic struc-
tures based on tied Gaussian mixture model and information criterion for con-
current sounds. In IEEE International Conference on Acoustics, Speech, and
Signal Processing, Montreal, Canada, 2004.
325. I. Kaminskyj and T. Czaszejko. Automatic recognition of isolated monophonic
musical instrument sounds using KNNC. Journal of Intelligent Information
Systems, 24:199-221, 2005.
326. A. Kapur, M. Benning, and G. Tzanetakis. Query-by-beat-boxing: Music re-
trieval for the DJ. In International Conference on Music Information Retrieval,
pp. 170-177, Barcelona, Spain, October 2004.
327. M. Karjalainen and T. Tolonen. Multi-pitch and periodicity analysis model for
sound separation and auditory scene analysis. In IEEE International Confer-
ence on Acoustics, Speech, and Signal Processing, Phoenix, USA, 1999.
 References 409

328. M. Karjalainen and T. Tolonen. Separation of speech signals using iterative

multipitch analysis and prediction. In 6th European Conf. Speech Communica-
tion and Technology, pp. 2187-2190, Budapest, Hungary, 1999.
329. M. Kartomi. On Concepts and Classification of Musical Instruments. The Uni-
versity of Chicago Press, 1990.
330. K. Kashino and H. Murase. A music stream segregation system based on adap-
tive multi-agents. In International Joint Conference on Artificial Intelligence,
pp. 1126-1136, Nagoya, Japan, 1997.
331. K. Kashino and H. Murase. A sound source identification system for ensem-
ble music based on template adaptation and music stream extraction. Speech
Communication, 27:337-349, 1999.
332. K. Kashino, K. Nakadai, T. Kinoshita, and H. Tanaka. Organisation of hier-
archical perceptual sounds: Music scene analysis with autonomous processing
modules and a quantitative information integration mechanism. In Interna-
tional Joint Conference on Artificial Intelligence, pp. 158-164, Montreal, Que-
bec, 1995.
333. K. Kashino and H. Tanaka. A sound source separation system with the abil-
ity of automatic tone modeling. In International Computer Music Conference,
pp. 248-255, Tokyo, Japan, 1993.
334. K. Kashino and H. Tanaka. A computational model of segregation of two fre-
quency components: Evaluation and integration of multiple cues. Electronics
and Communications in Japan, 77(7):35-47, 1994.
335. H. Katayose and S. Inokuchi. An intelligent transcription system. In First
International Conference on Music Perception and Cognition, pp. 95-98, 1989.
336. S.M. Katz. Estimation of probabilities from sparse data for the language model
component of a speech recogniser. IEEE Transactions on Acoustics, Speech,
and Signal Processing, 35(3):400-401, March 1987.
337. H. Kawahara, Y. Hirachi, M. Morise, and H. Banno. Procedure "senza vi-
brato" : A key component for morphing singing. In ISC A Tutorial and Research
Workshop on Statistical and Perceptual Audio Processing, Jeju, Korea, October
2004.
338. H. Kawahara, H. Katayose, A. de Cheveigne, and R.D. Patterson. Fixed point
analysis of frequency to instantaneous frequency mapping for accurate estima-
tion of FO and periodicity. In 6th European Conf. Speech Communication and
Technology, pp. 2781-2784, Budapest, Hungary, 1999.
339. S.M. Kay. Fundamentals of Statistical Signal Processing: Estimation Theory.
Prentice Hall, 1993.
340. R.A. Kendall. The role of acoustic signal partitions in listener categorization
of musical phrases. Music Perception, 4:185-214, 1986.
341. H.-G. Kim, J.J. Burred, and T. Sikora. How efficient is MPEG-7 for general
sound recognition? In 25th International Audio Engineering Society Conference
Metadata for Audio, London, UK, 2004.
342. H.-G. Kim, N. Moreau, and T. Sikora. MPEG-7 Audio and Beyond: Audio
Content Indexing and Retrieval. Wiley, 2005.
343. T. Kinoshita, S. Sakai, and H. Tanaka. Musical sound source identification
based on frequency component adaptation. In International Joint Conference
on Artificial Intelligence, pp. 18-24, Stockholm, Sweden, 1999.
344. G. Kitagawa. Monte Carlo filter and smoother for non-Gaussian nonlinear state
space models. Journal of Computational and Graphical Statistics, 5(1): 1-25,
1996.
410 References

345. T. Kitahara, M. Goto, and H.G. Okuno. Musical instrument identification

based on FO-dependent multivariate normal distribution. In IEEE Interna-
tional Conference on Acoustics, Speech, and Signal Processing, pp. 421-424,
Hong Kong, China, 2003.
346. T. Kitahara, M. Goto, and H.G. Okuno. Category-level identificatioon of non-
registered musical instrument sounds. In IEEE International Conference on
Acoustics, Speech, and Signal Processing, pp. 253-256, Montreal, Canada, 2004.
347. A. Klapuri. Sound onset detection by applying psychoacoustic knowledge. In
IEEE International Conference on Acoustics, Speech, and Signal Processing,
Phoenix, USA, 1999.
348. A. Klapuri. Musical meter estimation and music transcription. In Cambridge
Music Processing Colloquium, pp. 40-45, 2003.
349. A. Klapuri, A Eronen, and J. Astola. Analysis of the meter of acoustic musical
signals. IEEE Transactions on Speech and Audio Processing, 14(1), 2006.
350. A. Klapuri, T. Virtanen, A. Eronen, and J. Seppanen. Automatic transcription
of musical recordings. In Workshop on Consistent and Reliable Acoustic Cues
for Sound Analysis, Aalborg, Denmark, 2001.
351. A.P. Klapuri. Multiple fundamental frequency estimation based on harmonicity
and spectral smoothness. IEEE Transactions on Speech and Audio Processing,
11(6):804-815, 2003.
352. A.P. Klapuri. Automatic music transcription as we know it today. Journal of
New Music Research, 33(3), 2004.
353. A.P. Klapuri. Signal Processing Methods for the Automatic Transcription of
Music. PhD thesis, Tampere University of Technology, 2004.
354. A.P. Klapuri. A perceptually motivated multiple-FO estimation method for
polyphonic music signals. In IEEE Workshop on Applications of Signal Process-
ing to Audio and Acoustics, New Paltz, USA, 2005.
355. W.B. Kleijn and K.K. Paliwal, editors. Speech Coding and Synthesis. Elsevier
Academic Press, Amsterdam, 1995.
356. T. Kohonen. Emergence of invariant-feature detectors in the adaptive-subspace
self-organizing map. Biological Cybernetics, 75:281-291, 1996.
357. B. Kostek. Application of learning algorithms to musical sound analysis. In
Audio Engineering Society 97th Convention, San Francisco, USA, 1994.
358. B. Kostek. Feature extraction methods for the intelligent processing of musical
sounds. In Audio Engineering Society 99th Convention, New York, NY, 1995.
359. B. Kostek. Soft computing-based recognition of musical sounds. In L. Polkowski
and A. Skowron, editors. Rough Sets in Knowledge Discovery, pp. 193-213.
Physica-Verlag, Heidelberg, 1998.
360. B. Kostek. Soft Computing in Acoustics: Applications of Neural Networks,
Fuzzy Logic and Rough Sets to Musical Acoustics. Physica-Verlag, Heidelberg,
1999.
361. B. Kostek. Application of soft computing to automatic music information re-
trieval. Journal of the American Society for Information Science and Technol-
ogy, 55:1108-1116, 2004.
362. B. Kostek. Musical instrument recognition and duet analysis employing mu-
sic information retrieval techniques. In Proceedings of the IEEE, Volume 92,
pp. 712-729, 2004.
363. B. Kostek and A. Czyzewski. An approach to the automatic classification of
musical sounds. In Audio Engineering Society 108th Convention, Paris, France,
2000.
 References 411

364. B. Kostek and A. Czyzewski. Representing musical instrument sounds for their
automatic classification. Journal of the Audio Engineering Society, 49:768-785,
2001.
365. B. Kostek, M. Dziubinski, and P. Zwan. Further developments of methods for
searching optimum musical and rhythmic feature vectors. In Audio Engineering
Society 21st International Conference, St. Petersburg, Russia, 2002.
366. B. Kostek and R. Krolikowski. Application of artificial neural networks to the
recognition of musical sounds. Archives of Acoustics, 22:27-50, 1997.
367. B. Kostek, P. Szczuko, P. Zwan, and P Dalka. Processing of musical data
employing rough sets and artificial neural networks. In Transactions on Rough
Sets III, pp. 112-133, 2005.
368. B. Kostek and A. Wieczorkowska. Study of parameter relations in musical
instrument patterns. In Audio Engineering Society 100th Convention, Copen-
hangen, Denmark, 1996.
369. B. Kostek and A. Wieczorkowska. Parametric representation of musical sounds.
Archives of Acoustics, 22:3-26, 1997.
370. B. Kostek and P. Zwan. Wavelet-based automatic recognition of musical instru-
ment classes. In 142nd Meeting of the Acoustical Society of America, Melville,
New York, 2001.
371. B. Kostek, P. Zwan, and M. Dziubinski. Statistical analysis of musical sound
features derived from wavelet representation. In Audio Engineering Society
112nd Convention, Munich, Germany, 2002.
372. J.R. Koza. Genetic Programming: On the Programming of Computers by Means
of Natural Selection. MIT Press, 1992.
373. K. Kreutz-Delgado, J.F. Murray, B.D. Rao, K. Engan, T. Lee, and T.J. Se-
jnowski. Dictionary learning algorithms for sparse representation. Neural Com-
putation, 15:349-396, 2003.
374. J. Krimphoff, S. McAdams, and S. Winsberg. Caracterisation du timbre des
sons complexes. II: Analyses acoustiques et quantification psychophysique.
Journal de Physique, 4:625-628, 1994.
375. A. Krishna and T. Sreenivas. Music instrument recognition: Prom isolated notes
to solo phrcises. In IEEE International Conference on Acoustics, Speech, and
Signal Processing, Montreal, Canada, 2004.
376. S. Krstulovic, R. Gribonval, P. Leveau, and L. Daudet. A comparison of two
extensions of the matching pursuit algorithm for the harmonic decomposition
of sounds. In IEEE Workshop on Applications of Signal Processing to Audio
and Acoustics, New Paltz, USA, 2005.
377. C. Krumhansl. Cognitive Foundations of Musical Pitch. Oxford University
Press, 1990.
378. C.L. Krumhansl. Why is musical timbre so hard to understand? In
S. Nielzenand and O. Olsson, editors. Structure and Perception of Electroa-
coustic Sound and Music, pp. 43-53. Elsevier Academic Press, Amsterdam,
1989.
379. L. Kuncheva. Combining Pattern Classifiers: Methods and Algorithms. Wiley,
2004.
380. N. Kunieda, T. Shimamura, and J. Suzuki. Robust method of measurement of
fundamental frequency by ACLOSautocorrelation of log spectrum. In IEEE
International Conference on Acoustics, Speech, and Signal Processing, pp. 232-
235, Atlanta, USA, 1996.
412 References

381. J. Tin-Yau Kwok. Moderating the outputs of support vector machine classifiers.
IEEE Transactions on Neural Networks, 10(5):1018-1031, September 1999.
382. T.I. Laakso, V. VaUmaki, M. Karjalainen, and U.K. Laine. Splitting the unit
delay: Tools for fractional delay filter design. IEEE Signal Processing Magazine,
13(l):30-60, 1996.
383. M. Lagrange, S. Marchand, and J.B. Rault. Using linear prediction to en-
hance the tracking of partials. In IEEE International Conference on Acoustics,
Speech, and Signal Processing, Montreal, Canada, 2004.
384. M. Lahat, R. Niederjohn, and D.A. Krubsack. Spectral autocorrelation method
for measurement of the fundamental frequency of noise-corrupted speech. IEEE
Transactions on Acoustics, Speech, and Signal Processing, 6:741-750, June
1987.
385. S. Lakatos. A common perceptual space for harmonic and percussive timbres.
Perception and Psychophysics, 62:1426-1439, 1994.
386. T.L. Lam. Beat Tracking. Master's thesis. Department of Engineering, Univer-
sity of Cambridge, 2003.
387. D. Lang and N. de Frietas. Beat tracking the graphical model way. In Neural
Information Processing Systems, Vancouver, Canada, 2004.
388. K. Lange. Numerical Analysis for Statisticians. Springer, New York, USA,
1999.
389. E.W. Large. Dynamic Representation of Musical Structure. PhD thesis, Ohio
State Univ., 1994.
390. E.W. Large. Beat tracking with a nonUnear oscillator. In International Joint
Conference on Artificial Intelligence, pp. 24-31, Stockholm, Sweden, 1995.
391. E.W. Large and J.F. Kolen. Resonance and the perception of musical meter.
Connection Science, 6(l):177-208, 1994.
392. J. Laroche. Estimating tempo, swing and beat locations in audio recordings. In
IEEE Workshop on Applications of Signal Processing to Audio and Acoustics,
pp. 135-138, 2001.
393. J. Laroche. Efficient tempo and beat tracking in audio recordings. Journal of
the Audio Engineering Society, 51(4):226-233, April 2003.
394. J. Laroche, Y. Stylianou, and E. Mouhnes. HNS: Speech modification based
on a harmonic + noise model. In IEEE International Conference on Acoustics,
Speech, and Signal Processing, Volume 2, pp. 550-553, Minneapolis, USA, April
1993.
395. H. Laurent and C. Doncarli. Stationarity index for abrupt changes detection
in the time-frequency plane. IEEE Signal Processing Letters, 5(2):43-45, 1998.
396. S. Lauritzen and D. Spiegelhalter. Local computations with probabilities on
graphical structures and their application to expert systems. Journal of the
Royal Statistical Society: Series B, 50(2): 157-224, 1988.
397. C.L. Lawson and R.J. Hanson. Solving Least Squares Problems. Prentice Hall,
Englewood CHffs, New Jersey, 1974.
398. C.S. Lee. The perception of metrical structure: Experimental evidence and
a model. In P. Howell, R. West, and I. Cross, editors, Representing Musical
Structure. Academic Press, London, 1991.
399. D.D. Lee and H.S. Seung. Learning the parts of objects by non-negative matrix
factorization. Nature, 401:788-791, October 1999.
400. D.D. Lee and H.S. Seung. Algorithms for non-negative matrix factorization. In
Neural Information Processing Systems, pp. 556-562, Denver, USA, 2001.
 References 413

401. T. Lee and R. Orglmeister. A contextual blind separation of delayed and con-
volved sources. In IEEE International Conference on Acoustics, Speech, and
Signal Processing, pp. 1199-1202, Munich, Germany, 1997.
402. M. Leman. Music and Schema Theory. Springer, Heidelberg, 1995.
403. P. Lepain. Polyphonic pitch extraction from musical signals. Journal of New
Music Research, 28(4):296-309, 1999.
404. F. Lerdahl and R. Jackendoff. A Generative Theory of Tonal Music. MIT Press,
1983.
405. M. Lesaffre, M. Leman, B. De Baets, and J.-P. Martens. Methodological con-
siderations concerning manual annotation of musical audio in function of al-
gorithm development. In International Conference on Music Information Re-
trieval, Barcelona, Spain, 2004.
406. M. Lesaffre, M. Leman, K. Tanghe, B. De Baets, H. De Meyer, and
J.P. Martens. User-dependent taxonomy of musical features as a conceptual
framework for musical audio-mining technology. In Stockholm Music Acoustics
Conference, Stockholm, Sweden, 2003.
407. V. Lesser, S.H. Nawab, I. Gallastegi, and F. Klassner. IPUS: An architecture for
integrated signal processing and signal interpretation in complex environments.
In 11th National Conference on Artificial Intelligence, pp. 249-255, 1993.
408. S. Levine. Audio Representations for Data Compression and Compressed Do-
main Processing. PhD thesis. Center for Computer Research in Music and
Acoustics, Stanford University, 1998.
409. J.C.R. Licklider. A duplex theory of pitch perception. Experientia, 7:128-133,
1951. Reproduced in Schubert, E.D. (Ed.): Psychological Acoustics (Benchmark
papers in acoustics, vol. 13), Stroudsburg, Pennsylvania: Dowden, hutchinson
&; Ross, Inc., pp. 155-160.
410. T.M. Little and F.J. Hills. Statistical Methods in Agricultural Research. Uni-
versity of CaUfornia Press, 1972.
411. B. Liu, Y. Wu, and Y. Li. Linear hidden Markov model for music information
retrieval based on humming. In IEEE International Conference on Acoustics,
Speech, and Signal Processing, Volume 5, pp. 533-536, Hong Kong, China,
2003.
412. M. Liu and C. Wan. Feature selection for automatic classification of musi-
cal instrument sounds. In ACM/IEEE Joint Conference on Digital Libraries,
Roanoke, VA, USA, 2001.
413. A. Livshin, G. Peeters, and X. Rodet. Studies and improvements in automatic
classification of musical sound samples. In International Computer Music Con-
ference, Singapore, 2001.
414. A. Livshin and X. Rodet. The importance of cross database evaluation in
sound classification. In International Conference on Music Information Re-
trieval, Baltimore, USA, 2003.
415. A. Livshin and X. Rodet. Instrument recognition beyond separate notes: In-
dexing continuous recordings. In International Computer Music Conference,
Miami, Florida, USA, 2004.
416. A. Livshin and X. Rodet. Musical instrument identification in continuous
recordings. In International Conference on Digital Audio Effects, Naples, Italy,
2004.
417. B. Logan and S. Chu. Music summarization using key phrases. In IEEE In-
ternational Conference on Acoustics, Speech, and Signal Processing, Volume 2,
pp. 749-752, Istanbul, Turkey, 2000.
414 References

418. H.C. Longuet-Higgins and C.S. Lee. The perception of musical rhythms. Per-
ception, 11 (2): 115-128, 1982.
419. M.A. Loureiro, H.B. de Paula, and H.C. Yehia. Timbre classification of a sin-
gle musical instrument. In International Conference on Music Information Re-
trieval, Barcelona, Spain, 2004.
420. L. Lu, M. Wang, and H.-J. Zhang. Repeating pattern discovery and structure
analysis from acoustic music data. In ACM SIGMM International Workshop
on Multimedia Information Retrieval, pp. 275-282, 2004.
421. L. Lu, H. You, and H.-J. Zhang. A new approach to query by humming in
music retrieval. In IEEE International Conference on Multimedia and Expo,
pp. 776-779, Tokyo, Japan, 2001.
422. R.F. Lyon. Computational models of neural auditory processing. In IEEE In-
ternational Conference on Acoustics, Speech, and Signal Processing, pp. 36.1.1-
36.1.4, San Diego, California, 1984.
423. A. Madevska-Bogdanova and D. Nikohc. A geometrical modification of SVM
outputs for pattern recognition. In 22nd International Conference on Informa-
tion Technology Interfaces, pp. 165-170, Pula, Croatia, June 2000.
424. A. Madevska-Bogdanova and D. Nikolic. A new approach of modifying SVM
outputs. In International Joint Conference on Neural Networks, Volume 6,
pp. 395-398, Como, Italy, July 2000.
425. R. Maher and J. Beauchamp. An investigation of vocal vibrato for synthesis.
Applied Acoustics, 30:219-245, 1990.
426. R.C. Maher. An Approach for the Separation of Voices in Composite Music
Signals. PhD thesis, Univ. of Illinois, Urbana, 1989.
427. R.C. Maher. Evaluation of a method for separating digitized duet signals. Jour-
nal of the Audio Engineering Society, 38(12):956-979, 1990.
428. R.C. Maher and J.W. Beauchamp. Fundamental frequency estimation of mu-
sical signals using a two-way mismatch procedure. Journal of the Acoustical
Society of America, 95(4):2254-2263, April 1994.
429. S. Mallat. A Wavelet Tour of Signal Processing. Academic Press, 1998.
430. S. Mallat and Z. Zhang. Matching pursuits with time-frequency dictionaries.
IEEE Transactions on Signal Processing, 41:3397-3415, 1993.
431. H.S. Malvar. Signal Processing with Lapped Transforms. Artech House, Nor-
wood, MA, 1992.
432. B.S. Manjunath, P. Salembier, and T. Sikora. Introduction to MPEG-7: Mul-
timedia Content Description Language. John Wiley & Sons, 2002.
433. M. Marolt. SONIC: Transcription of polyphonic piano music with neural net-
works. In MOSART Workshop on Current Research Directions in Computer
Music, Barcelona, Spain, November 2001.
434. M. Marolt. A connectionist approach to transcription of polyphonic piano mu-
sic. IEEE Transactions on Multimedia, 6(3):439-449, 2004. URL: Igm.fri.uni-
Ij.si/SONIC.
435. M. Marolt. Gaussian mixture models for extraction of melodic lines from au-
dio recordings. In International Conference on Music Information Retrieval,
Barcelona, Spain, October 2004.
436. M. Marolt. On finding melodic lines in audio recordings. In International Con-
ference on Digital Audio Effects, Naples, Italy, 2004.
437. J. Marques and P.J. Moreno. A study of musical instrument classification using
Gaussian mixture models and support vector machines. Technical Report CRL
99/4, Compaq, 1999.
 References 415

438. D. Marr. Vision. W.H. Freeman and Company, 1982.

439. K.D. Martin. Automatic transcription of simple polyphonic music: Robust
front end processing. Technical Report 399, MIT Media Laboratory Percep-
tual Computing Section, 1996.
440. K.D. Martin. A blackboard system for automatic transcription of simple poly-
phonic music. Technical Report 385, MIT Media Laboratory Perceptual Com-
puting Section, 1996.
441. K.D. Martin. Toward automatic sound source recognition: Identifying musical
instruments. In Computational Hearing Conference of NATO Advanced Study
Institute, Tuscany, Italy, 1998.
442. K.D. Martin. Sound-Source Recognition: A Theory and Computational Model.
PhD thesis, Massachusetts Institute of Technology, 1999.
443. K.D. Martin and Y.E. Kim. Musical instrument identification: A pattern-
recognition approach. In 136th meeting of the Acoustical Society of America,
Norfolk, Virginia, USA, 1998.
444. P. Masri and A. Bateman. Improved modelling of attack transients in music
analysis-resynthesis. In International Computer Music Conference, pp. 100-
103, Hong Kong, China, August 1996.
445. O. Mayor. An adaptive real-time beat tracking system for polyphonic pieces of
audio using multiple hypotheses. In M OSART Workshop on Current Research
Directions in Computer Music, 2001.
446. S. McAdams, J.W. Beauchamp, and S. Meneguzzi. Discrimination of musical
instrument sounds resynthesized with simplified spectrotemporal parameters.
Journal of the Acoustical Society of America, 105:882-897, 1999.
447. S. McAdams and A.S. Bregman. Hearing musical streams. Computer Music
Journal, 3(4):26-43, 1979.
448. S. McAdams, S. Winsberg, G. de Soete, and J. Krimphoff. Perceptual scaling
of synthesized musical timbres: Common dimensions, specificities, and latent
subject classes. Psychological Research, 58:177-192, 1995.
449. R.J. McAulay and Th.F. Quatieri. Speech analysis/synthesis based on a si-
nusoidal representation. IEEE Transactions on Acoustics, Speech, and Signal
Processing, 34:744-754, 1986.
450. J.D. McAuley. Perception of Time as Phase: Towards an Adaptive-Oscillator
Model of Rhythmic Pattern Processing. PhD thesis, Computer Science and
Cognitive Science, Univ. Indiana, July 1995.
451. M.E. Mclntyre, R.T. Schumacher, and J. Woodhouse. On the oscillations of
musical instruments. Journal of the Acoustical Society of America, 74(5): 1325-
1345, 1983.
452. M.F. McKinney and D. Moelants. Extracting the perceptual tempo from music.
In International Conference on Music Information Retrieval, Barcelona, Spain,
October 2004.
453. G.J. McLachlan. Discriminant Analysis and Statistical Pattern Recognition.
John Wiley & Sons, 1992.
454. R.J. McNab, L.A. Smith, and I.H. Witten. Signal processing for melody tran-
scription. In 19th Australasian Computer Science Conference, Melbourne, Aus-
traha, February 1996.
455. R.J. McNab, L.A. Smith, I.H. Witten, and C.L. Henderson. Tune retrieval
in the multimedia library. Multimedia Tools and Applications, 10(2): 113-132,
April 2000.
416 References

456. R. Meddis. Simulation of mechanical to neural transduction in the auditory

receptor. Journal of the Acoustical Society of America, 79(3):702-711, 1986.
457. R. Meddis and M.J. Hewitt. Virtual pitch and phase sensitivity of a computer
model of the auditory periphery. I: Pitch identification. II: Phase sensitivity.
Journal of the Acoustical Society of America, 89(6):2866-2894, 1991.
458. R. Meddis and M.J. Hewitt. Modeling the identification of concurrent vowels
with different fundamental frequencies. Journal of the Acoustical Society of
America, 91(l):233-245, 1992.
459. R. Meddis and L. O'Mard. A unitary model of pitch perception. Journal of the
Acoustical Society of America, 102(3):1811-1820, 1997.
460. D.K. Mellinger. Event Formation and Separation of Musical Sound. PhD thesis,
Stanford University, Stanford, USA, 1991.
461. Y. Meron and K. Hirose. Synthesis of vibrato singing. In IEEE International
Conference on Acoustics, Speech, and Signal Processing, Volume 2, pp. 745-
748, Istanbul, Turkey, 2000.
462. The MIDI Manufacturers Association. The Complete MIDI 1.0 Detailed Spec-
ification, second edition, 1996. Website: www.midi.org.
463. I. Mierswa and K. Morik. Automatic feature extraction for classifying audio
data. Machine Learning, 58:127-149, 2005.
464. B.O. Miller, D.L. Scarborough, and J.A. Jones. On the perception of meter. In
M. Balaben, K. Ebeioglu, and O. Laske, editors. Understanding Music With
AI, chapter 18, pp. 428-47. MIT Press, 1992.
465. MIREX: Annual Music Information Retrieval Evaluation eXchange, 2005.
URL: www.music-ir.org/ mirexwiki/ index.php/ MIREX_2005.
466. O.M.E. Mitchell, C.A. Ross, and G.H. Yates. Signal processing for a cocktail
party effect. Journal of the Acoustical Society of America, 50(2):656-660, 1971.
467. T.M. Mitchell. Machine Learning. McGraw-Hill, 1997.
468. S. Molla and B. Torresani. Determining local transientness of audio signals.
IEEE Signal Processing Letters, ll(7):625-628, 2004.
469. S. Molla and B. Torresani. A hybrid scheme for encoding audio signals using
hidden Markov models of waveforms. Applied and Computational Harmonic
Analysis, 2005. (in press)
470. B. Mont-Reynaud. Problem-solving strategies in a music transcription system.
In 9th International Joint Conference on Artificial Intelligence, pp. 916-918,
Los Angeles, USA, 1985.
471. T.K. Moon. The expectation-maximization algorithm. IEEE Signal Processing
Magazine, 13(6):47-60, November 1996.
472. T.K. Moon and W.C. Stirfing. Mathematical Methods and Algorithms for Signal
Processing. Prentice Hall, Upper Saddle River, NJ, 2000.
473. B.C.J. Moore. Frequency analysis and masking. In Moore [474], pp. 161-205.
474. B.C.J. Moore, editor. HearingHandbook of Perception and Cognition. Acad-
emic Press, San Diego, California, 2nd edition, 1995.
475. B.C.J. Moore. An Introduction to the Psychology of Hearing. Academic Press,
4th edition, 1997.
476. B.C.J. Moore, B.R. Glasberg, and R.W. Peters. Thresholds for hearing mis-
tuned partials as separate tones in harmonic complexes. Journal of the Acousti-
cal Society of America, 80(2):479-483, 1986.
477. J.A. Moorer. On the Segmentation and Analysis of Continuous Musical Sound
by Digital Computer. PhD thesis, Dept. of Music, Stanford University, 1975.
Distributed as Dept. of Music report No. STAN-M-3.
 References 417

478. J.A. Moorer. On the transcription of musical sound by computer. Computer

Music Journal, l(4):32-38, 1977.
479. T. Nakano, J. Ogata, M. Goto, and Y. Hiraga. A drum pattern retrieval method
by voice percussion. In International Conference on Music Information Re-
trieval, pp. 550-553, Barcelona, Spain, October 2004.
480. T. Nakatani. Computational Auditory Scene Analysis Based on Residue-Driven
Architecture and Its Application to Mixed Speech Recognition. PhD thesis, Ky-
oto University, Kyoto, Japan, 2002.
481. T. Nakatani, G.H. Okuno, and T. Kawabata. Residue-driven architecture for
computational auditory scene analysis. In International Joint Conference on
Artificial Intelligence, pp. 165-172, Montreal, Quebec, 1995.
482. A. Nehorai and B. Porat. Adaptive comb filtering for harmonic signal enhance-
ment. IEEE Transactions on Acoustics, Speech, and Signal Processing, 34(5),
1986.
483. T. Niihara and S. Inokuchi. Transcription of sung song. In IEEE International
Conference on Acoustics, Speech, and Signal Processing, pp. 1277-1280, Tokyo,
Japan, 1986.
484. T. Nishimura, H. Hashiguchi, J. Takita, J. Xin Zhang, M. Goto, and R. Oka.
Music signal spotting retrieval by a humming query using start frame fea-
ture dependent continuous dynamic programming. In 2nd Annual Interna-
tional Symposium on Music Information Retrieval, pp. 211-218, Bloomington,
Indiana, USA, October 2001.
485. H. Okuno and D Rosenthal, editors. Computational Auditory Scene Analysis.
Lawrence Erlbaum Associates, Pubhshers, 1997.
486. B.A. Olshausen and D.F. Field. Sparse coding with an overcomplete basis set:
A strategy employed by vl? Vision Research, 37:3311-3325, 1997.
487. F. Opolko and J. Wapnick. McGill University Master Samples. Technical re-
port, McGill University, Montreal, Canada, 1987.
488. I.F.O Orife. Riddim: A Rhythm Analysis and Decomposition Tool Based on
Independent Subspace Analysis. Master's thesis, Dartmouth College, Hanover,
New Hampshire, USA, May 2001.
489. N. Orio and M.S. Sette. A HMM-based pitch tracker for audio queries. In Inter-
national Conference on Music Information Retrieval, pp. 249-250, Baltimore,
USA, 2003.
490. J. Osborne. Notes on the use of data transformations. Practical Assessment,
Research and Evaluation, 8(6), 2002.
491. N. Otsu. A threshold selection method from gray-level histograms. IEEE Trans-
actions on Systems, Man, and Cybernetics, SMC-9(l):62-66, 1979.
492. P. Paatero and U. Tapper. Positive matrix factorization: A non-negative fac-
tor model with optimal utilization of error estimates of data values. Environ-
metrics, 5:111-126, 1994.
493. R.P. Paiva, T. Mendes, and A. Cardoso. An auditory model based approach
for melody detection in polyphonic musical recordings. In International Sym-
posium on Computer Music Modeling and Retrieval, May 2004.
494. R.P. Paiva, T. Mendes, and A. Cardoso. A methodology for detection of melody
in polyphonic musical signals. In Audio Engineering Society 116th Convention,
Berlin, Germany, May 2004.
495. T.H. Park. Towards Automatic Musical Instrument Timbre Recognition. PhD
thesis, Princeton University, 2004.
418 References

496. T.H. Park and P. Cook. Nearest centroid error clustering for radial/elliptical
basis function neural networks in timbre classification. In International Com-
puter Music Conference, pp. 833-866, Barcelona, Spain, 1998.
497. R. Parncutt. A perceptual model of pulse salience and metrical accent in mu-
sical rhythms. Music Perception, ll(4):409-464, 1994.
498. T.W. Parsons. Separation of speech from interfering speech by means of har-
monic selection. Journal of the Acoustical Society of America, 60(4), 1976.
499. R.D. Patterson. Auditory filter shapes derived with noise stimuli. Journal of
the Acoustical Society of America, 59(3):640-654, 1976.
500. R.D. Patterson. Auditory images: How complex sounds are represented in the
auditory system. Journal of the Acoustical Society of Japan (E), 21(4):183-190,
2000.
501. R.D. Patterson and M.H. Allerhand. Time-domain modeling of peripheral audi-
tory processing: A modular architecture and a software platform. Journal of the
Acoustical Society of America, 98(4): 1890-1894, 1995. URL: http://www.mrc-
cbu.cam.ac.uk/cnbh/web2002/ bodyframes/AIM.htm.
502. R.D. Patterson and J. Holdsworth. A functional model of neural activity pat-
terns and auditory images. In Ainsworth [13], pp. 551-567.
503. J. Paulus and A. Klapuri. Measuring the similarity of rhythmic patterns. In
International Conference on Music Information Retrieval, Paris, France, 2002.
504. J. Paulus and A. Klapuri. Model-based event labeling in the transcription of
percussive audio signals. In M. Davies, editor. International Conference on
Digital Audio Effects, pp. 73-77, London, UK, September 2003.
505. J. Paulus and T. Virtanen. Drum transcription with non-negative spectrogram
factorisation. In European Signal Processing Conference, Antalya, Turkey, Sep-
tember 2005.
506. J.K. Paulus and A.P. Klapuri. Conventional and periodic N-grams in the tran-
scription of drum sequences. In IEEE International Conference on Multimedia
and Expo, Volume 2, pp. 737-740, Baltimore, Maryland, USA, July 2003.
507. Steffen Pauws. CubyHum: A fully operational query by humming system. In
International Conference on Music Information Retrieval, pp. 187-196, Paris,
France, October 2002.
508. Z. Pawlak. Rough set elements. In L. Polkowski and A. Skowron, editors. Rough
Sets in Knowledge Discovery, pp. 10-30. Physica-Verlag, Heidelberg, 1998.
509. J. Pearl. Probabilistic Reasoning in Intelligent Systems: Networks of Plausible
Inference. Morgan Kaufmann Pubhshers, 1988.
510. G. Peeters. Automatic classification of large musical instrument databases us-
ing hierarchical classifiers with inertia ratio maximization. In Audio Engineer-
ing Society 115th Convention, New York, NY, USA, 2003.
511. G. Peeters. A large set of audio features for sound description (similarity
and classification) in the CUIDADO project. Technical report, IRC AM, Paris,
France, April 2004.
512. G. Peeters, A. La Burthe, and X. Rodet. Toward automatic music audio sum-
mary generation from signal analysis. In International Conference on Music
Information Retrieval, pp. 94-100, Paris, France, October 2002.
513. G. Peeters, S. McAdams, and P. Herrera. Instrument sound description in the
context of MPEG-7. In International Computer Music Conference, pp. 166-
169, Berhn, Germany, 2000.
 References 419

514. G. Peeters and X. Rodet. Automatically selecting signal descriptors for sound
classification. In International Computer Music Conference, Goteborg, Sweden,
September 2002.
515. G. Peeters and X. Rodet. Hierarchical Gaussian tree with inertia ratio maxi-
mization for the classification of large musical instrument databases. In Inter-
national Conference on Digital Audio Effects, London, UK, 2003.
516. G. Peeters and X. Rodet. Signal-based music structure discovery for music au-
dio summary generation. In International Computer Music Conference, pp. 15-
22, Singapore, 2003.
517. I. Peretz. Music perception and recognition. In B. Rapp, editor. The Handbook
of Cognitive Neuropsychology, pp. 519-540. Hove: Psychology Press, 2001.
518. I. Peretz and M. Coltheart. Modularity of music processing. Nature Neuro-
science, 6(7), 2003.
519. G. Peterschmitt, E. Gomez, and P. Herrera. Pitch-based solo location. In
MOSART Workshop on Current Research Directions in Computer Music,
Barcelona, Spain, 2001.
520. M. Piszczalski. A Computational Model of Music Transcription. PhD thesis,
Univ. of Michigan, Ann Arbor, 1986.
521. M. Piszczalski and B.A. Galler. Automatic music transcription. Computer Mu-
sic Journal, 1(4):24-31, 1977.
522. C.J. Plack, A.J. Oxenham, R.R. Fay and A.N. Popper, editors. Pitch. Springer,
New York, 2005.
523. C.J. Plack and R.P. Carlyon. Loudness perception and intensity coding. In
Moore [474], pp. 123-160.
524. J.C. Piatt. Probabilistic outputs for support vector machines and comparisons
to regularized likelihood methods. In A.J. Smola, P. Bartlett, B. Scholkopf,
and D. Schuurmans, editors. Advances in Large Margin Classifiers. MIT Press,
1999.
525. M.D. Plumbley. Conditions for non-negative independent component analysis.
IEEE Signal Processing Letters, 9(6), 2002.
526. M.D. Plumbley and E. Oja. A 'non-negative PC A' algorithm for indepen-
dent component analysis. IEEE Transactions on Neural Networks, 15(l):66-67,
2004.
527. H.F. Pollard and E.V. Janson. A tristimulus method for the specification of
musical timbre. Acustica, 51:162-171, 1982.
528. E. PoUastri. A pitch tracking system dedicated to process singing voice for
musical retrieval. In IEEE International Conference on Multimedia and Expo,
Volume 1, pp. 341-344, Lusanne, Switzerland, 2002.
529. D-J. Povel and P. Essens. Perception of musical patterns. Music Perception,
2(4):411-440, Summer 1985.
530. E. Prame. Vibrato extent and intonation in professional Western lyric singing.
Journal of the Acoustical Society of America, 102(1):616-621, July 1997.
531. W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery. Numerical
Recipes in C/C-h-h; The Art of Scientific Computing. Cambridge University
Press, Cambridge, UK, 2002.
532. H. Purnhagen and N. Meine. HILN: The MPEG-4 parametric audio coding
tools. In the IEEE International Symposium on Circuits and Systems (ISCAS
2000), Geneva, Switzerland, 2000.
533. J.R. Quinlan. CJi.-5: Programs for Machine Learning. Morgan Kaufmann Pub-
lishers, San Francisco, CA, USA, 1993.
420 References

534. L.R. Rabiner. A tutorial on hidden Markov models and selected applications
in speech recognition. Proc. of IEEE, 77(2):257-289, February 1989.
535. L.R. Rabiner, M.J. Cheng, A.E. Rosenberg, and C.A. McGonegal. A compar-
ative performance study of several pitch detection algorithms. IEEE Transac-
tions on Acoustics, Speech, and Signal Processing, 24(5):399-418, 1976.
536. L.R. Rabiner and B.-H. Juang. Fundamentals of Speech Recognition. Prentice
Hall, New Jersey, 1993.
537. C. Raphael. Automated rhythm transcription. In 2nd Annual International
Symposium on Music Information Retrieval, Bloomington, Indiana, USA, 2001.
538. C. Raphael. A probabilistic expert system for automatic musical accompani-
ment. Journal of Computational and Graphical Statistics, 10(3):486-512, 2001.
539. M. Reyes-Gomez, N. Jojic, and D. Ellis. Deformable spectrograms. In 10th
International Workshop on Artificial Intelligence and Statistics, pp. 285-292,
Barbados, 2005.
540. E. Rich and K. Knight. Artificial Intelligence. McGraw-Hill, New York, 1991.
541. S. Richardson and P.J. Green. On Bayesian analysis of mixtures with an un-
known number of components. Journal of the Royal Statistical Society: Series
B, 59(4):731-792, 1997.
542. C. Roads. Research in music and artificial intelligence. ACM Computing Sur-
veys, 17(2): 163-190, 1985.
543. C. Roads. The Computer Music Tutorial. MIT Press, Cambridge, USA, 1996.
544. C.P. Robert and G. Casella. Monte Carlo Statistical Methods. Springer, New
York, USA, 2000.
545. A. Ron and Z. Shen. Frames and stable bases for shift-invariant subspaces of
L^{R'^). Canadian Journal of Mathematics, 47:1051-1094, 1995.
546. D. Rosenthal. Emulation of human rhythm perception. Computer Music Jour-
nal, 16(l):64-72, Spring 1992.
547. D.F. Rosenthal. Machine Rhythm: Computer Ennulation of Human Rhythm
Perception. PhD thesis, Massachusetts Institute of Technology, 1992.
548. J. Rosenthal. A First Look at Rigorous Probability Theory. World Scientific
Publishing Company, 2000.
549. M.J. Ross, H.L. Shaffer, A. Cohen, R. Freudberg, and H.J. Manley. Av-
erage magnitude difference function pitch extractor. IEEE Transactions on
Acoustics, Speech, and Signal Processing, 22:353-362, 1974.
550. T.D. Rossing. The Science of Sound. Addison Wesley, second edition, 1990.
551. T.D. Rossing. Science of Percussion Instruments. World Scientific Publishing
Company, 2000.
552. C. Rover, F. Klefenz, and C. Weihs. Identification of Musical Instruments by
Means of the Hough-Transformation. Springer, 2005.
553. R. Rowe. Machine Musicianship. MIT Press, Cambridge, Massachusetts, 2001.
554. S. Roweis. One microphone source separation. In T.K. Leen, T.G. Dietterich,
and V. Tresp, editors, Neural Information Processing Systems, pp. 793-799,
Denver, USA, 2000.
555. D.E. Rumelhart, G.E. Hinton, and R.J. Williams. Learning internal representa-
tions by error propagation. In D.E. Rumelhart, J.L. McClelland, and the PDP
Research Group, editors. Parallel Distributed Processing: Explorations in the
Microstructure of Cognition, pp. 318-362. MIT Press, Cambridge, MA, 1986.
556. S. Riiping. A simple method for estimating conditional probabilities for SVMs.
Technical report, CS Department, Dortmund University, Dortmund, Germany,
December 2004.
 References 421

557. M. Ryynanen. Probabilistic Modelling of Note Events in the Transcription

of Monophonic Melodies. Master's thesis, Tampere University of Technology,
March 2004.
558. M. Ryynanen and A. Klapuri. ModelUng of note events for singing transcrip-
tion. In ISCA Tutorial and Research Workshop on Statistical and Perceptual
Audio Processing, Jeju, Korea, October 2004.
559. M. Ryynanen and A. Klapuri. Polyphonic music transcription using note event
modeling. In IEEE Workshop on Applications of Signal Processing to Audio
and Acoustics, New Paltz, USA, 2005.
560. G.J. Sandell and W.L. Martens. Perceptual evaluation of principal-component-
based synthesis of musical timbres. Journal of the Acoustical Society of Amer-
ica, 43:1013-1028, 1995.
561. V. Sand void, F. Gouyon, and P. Herrera. Percussion classification in polyphonic
audio recordings using localized sound models. In International Conference on
Music Information Retrieval, Barcelona, Spain, October 2004.
562. T.J. Santner, B.J. Williams, and W.I. Notz. The Design and Analysis of Com-
puter Experiments. Berlin: Springer Verlag, 2003.
563. L.K. Saul, F. Sha, and D.D. Lee. Statistical signal processing with nonnega-
tivity constraints. In Euro Speech, Geneva, Switzerland, 2003.
564. E.D. Scheirer. Tempo and beat analysis of acoustical musical signals. Journal
of the Acoustical Society of America, 103(1):588-601, January 1998.
565. E.D. Scheirer. Music-Listening Systems. PhD thesis, Massachusetts Institute
of Technology, 2000.
566. E.D. Scheirer. About this business of metadata. In International Conference
on Music Information Retrieval, pp. 252-254, 2002.
567. W.A. Schloss. On the Automatic Transcription of Percussive Music - from
Acoustic Signal to High-Level Analysis. PhD thesis. Center for Computer Re-
search in Music and Acoustics, Stanford University, Stanford, California, USA,
May 1985.
568. B. Scholkopf and A. Smola. Learning with Kernels. MIT Press, Cambridge,
USA, 2002.
569. E. Schuijers, W. Oomen, B. den Brinker, and J. Breebaart. Advances in para-
metric coding for high quality audio. In Audio Engineering Society 114th Con-
vention, Amsterdam, Netherlands, 2003.
570. C.E. Seashore. Psychology of Music. Dover Publications, 1967.
571. E. Selfridge-Field. Beyond MIDI: The Handbook of Musical Codes. MIT Press,
Cambridge, Massachusetts, 1997.
572. J. Seppanen. Computational Models of Musical Meter Estimation. Master's
thesis, Tampere University of Technology, Tampere, Finland, November 2001.
573. J. Seppanen. Tatum grid analysis of musical signals. In IEEE Workshop on
Applications of Signal Processing to Audio and Acoustics, pp. 131-134, New
Paltz, USA, 2001.
574. X. Serra. A System for Sound Analysis/Transformation/Synthesis Based on a
Deterministic Plus Stochastic Decomposition. PhD thesis, Stanford University,
1990.
575. X. Serra. Musical sound modeling with sinusoids plus noise. In C. Roads,
S. Pope, A. Picialli, and G. De Poli, editors. Musical Signal Processing. Swets
& Zeitlinger, 1997.
422 References

576. X. Serra and J.O. Smith. Spectral modeling synthesis: A sound analy-
sis/synthesis system based on a deterministic plus stochastic decomposition.
Computer Music Journal, 14(4): 12-24, Winter 1990.
577. W.A. Sethares, R.D. Morris, and J.C. Sethares. Beat tracking of musical per-
formances using low-level audio features. IEEE Transactions on Speech and
Audio Processing, 13(2): 1063-1076, 2005.
578. W.A. Sethares and T.A. Staley. Meter and periodicity in musical performance.
Journal of New Music Research, 30(2), June 2001.
579. F. Sha and L.K. Saul. Real-time pitch determination of one or more voices by
nonnegative matrix factorization. In Neural Information Processing Systems,
Vancouver, Canada, 2004.
580. G. Shafer. A Mathematical Theory of Evidence. Princeton University Press,
1976.
581. R.V. Shannon, F.-G. Zeng, V. Kamath, J. Wygonski, and M. Ekelid. Speech
recognition with primarily temporal cues. Science, 270(5234):303-304, 1995.
582. J.M. Shapiro. Embedded image coding using zerotrees of wavelet coefficients.
IEEE Transactions on Signal Processing, 41(12):3445-3462, 1993.
583. A. Sheh and D.P.W. Ellis. Chord segmentation and recognition using EM-
trained hidden Markov models. In International Conference on Music Infor-
mation Retrieval, pp. 183-189, Baltimore, USA, October 2003.
584. R.N. Shepard. Circularity in judgments of relative pitch. Journal of the Acousti-
cal Society of America, 36(12):2346-2353, 1964.
585. J. Shifrin, B. Pardo, C. Meek, and W. Birmingham. HMM-based musical query
retrieval. In ACM/IEEE Joint Conference on Digital Libraries, pp. 295-300,
Oregon, USA, 2002.
586. H. Shih, S.S. Narayanan, and C.-C.J. Kuo. A statistical multidimensional hum-
ming transcription using phone level hidden Markov models for query by hum-
ming systems. In IEEE International Conference on Multimedia and Expo,
Volume 1, pp. 61-64, Baltimore, Maryland, USA, 2003.
587. H.-H. Shih, S.S. Narayanan, and C.-C.J. Kuo. An HMM-based approach to
humming transcription. In IEEE International Conference on Multimedia and
Expo, Lusanne, Switzerland, 2002.
588. H.-H. Shih, S.S. Narayanan, and C.-C.J. Kuo. Multidimensional humming
transcription using a statistical approach for query by humming systems. In
IEEE International Conference on Acoustics, Speech, and Signal Processing,
Volume 5, pp. 541-544, Hong Kong, China, 2003.
589. J.I. Shonle and K.E. Horan. The pitch of vibrato tones. Journal of the Acousti-
cal Society of America, 67(l):246-252, January 1980.
590. J. Sillanpaa, A. Klapuri, J. Seppanen, and T. Virtanen. Recognition of acoustic
noise mixtures by combined bottom-up and top-down processing. In European
Signal Processing Conference, 2000.
591. M. Slaney. An efficient implementation of the Patterson Holdsworth audi-
tory filter bank. Technical Report 35, Perception Group, Advanced Technology
Group, Apple Computer, 1993.
592. M. Slaney. A critique of pure audition. In International Joint Conference on
Artificial Intelligence, pp. 13-18, Montreal, Quebec, 1995.
593. M. Slaney. Mixtures of probability experts for audio retrieval and indexing.
In IEEE International Conference on Multimedia and Expo, Lusanne, Switzer-
land, 2002.
 References 423

594. M. Slaney and R.F. Lyon. A perceptual pitch detector. In IEEE International
Conference on Acoustics, Speech, and Signal Processing, pp. 357-360, Albu-
querque, New Mexico, 1990.
595. M. Slaney and R.F. Lyon. On the importance of timea temporal represen-
tation of sound. In M. Cooke, S. Beet, and M. Crawford, editors, Visual Rep-
resentations of Speech Signals, pp. 95-116. John Wiley Sz Sons, 1993.
596. D. Sleator and D. Temper ley. Melisma music analyser code., 2001.
http://www.link.cs.cmu.edu/music-analysis/.
597. D. Slezak, R Synak, A. Wieczorkowska, and J. Wroblewski. KDD-based ap-
proach to musical instrument sound recognition. In M.-S. Hacid, Z.W. Ras,
D.A. Zighed, and Y. Kodratoff, editors. International Symposium on Method-
ologies for Intelligent Systems, Volume 2366 of Lecture Notes in Artificial In-
telligence, pp. 28-36. Springer, 2002.
598. P. Smaragdis. Redundancy Reduction for Computational Audition, a Unifying
Approach. PhD thesis, Massachusetts Institute of Technology, 2001.
599. P. Smaragdis. Discovering auditory objects through non-negativity constraints.
In ISCA Tutorial and Research Workshop on Statistical and Perceptual Audio
Processing, Jeju, Korea, 2004.
600. P. Smaragdis and J.C. Brown. Non-negative matrix factorization for poly-
phonic music transcription. In IEEE Workshop on Applications of Signal
Processing to Audio and Acoustics, New Paltz, USA, 2003.
601. L.M. Smith and P. Kovesi. A continuous time-frequency approach to repre-
senting rhythmic strata. In 4ih Int. Conf. on Music Perception and Cognition,
Montreal, Canada, 1996.
602. B. Snyder. Music and Memory. MIT Press, Cambridge, Massachusetts, 2000.
603. J. Song, S.Y. Bae, and K. Yoon. Mid-level music melody representation of
polyphonic audio for query-by-humming system. In International Conference
on Music Information Retrieval, pp. 133-139, Paris, France, October 2002.
604. T. Sonoda, M. Goto, and Y. Muraoka. A WWW-based melody retrieval system.
In International Computer Music Conference, pp. 349-352, Michigan, USA,
1998.
605. T. Sonoda, T. Ikenaga, K. Shimizu, and Y. Muraoka. The design method of a
melody retrieval system on parallelized computers. In International Conference
on Web Delivering of Music, pp. 66-73, Darmstadt, Germany, December 2002.
606. A. Srinivasan, D. Sullivan, and I. Fujinaga. Recognition of isolated instrument
tones by conservatory students. In 7th Int. Conf. on Music Perception and
Cognition, pp. 720-723, Sydney, Austraha, 2002.
607. M.J. Steedman. The perception of musical rhythm and metre. Perception,
6(5):555-569, 1977.
608. D. Van Steelant, K. Tanghe, S. Degroeve, M. Baets, B. De Leman, and
J.-P. Martens. Classification of percussive sounds using support vector ma-
chines. In Machine Learning Conference of Belgium and The Netherlands,
Brussels, Belgium, January 2004.
609. A. Sterian, M.H. Simoni, and G.H. Wakefield. Model-based musical transcrip-
tion. In International Computer Music Conference, Beijing, China, 1999.
610. A. Sterian and G.H. Wakefield. Music transcription systems: From sound to
symbol. In Workshop on AI and Music, 2000.
611. A.D. Sterian. Model-Based Segmentation of Time-Frequency Images for Mu-
sical Transcription. PhD thesis, MusEn Project, University of Michigan, Ann
Arbor, 1999.
424 References

612. J.V. Stone, J. Porrill, C. Buchel, and K. Priston. Spatial, temporal, and spa-
tiotemporal independent component analysis of fMRI data. In the 18th Leeds
Statistical Research Workshop on Spatial-Temporal Modelling and its Applica-
tions, 1999.
613. J. Sundberg. The Science of the Singing Voice. Northern Illinois University
Press, 1987.
614. J. Sundberg. The perception of singing. In D. Deutsch, editor, The Psychology
of Music, pp. 171-214. Academic Press, 1999.
615. P. Szczuko, P. Dalka, M. Dabrowski, and B. Kostek. MPEG-7-based low-level
descriptor effectiveness in the automatic musical sound classification. In Audio
Engineering Society 116th Convention, Berlin, Germany, 2004.
616. J. Tabrikian, S. Dubnov, and Y. Dickalov. Maximum a posteriori probability
pitch tracking in noisy environments using harmonic model. IEEE Transactions
on Speech and Audio Processing, 12(l):76-87, 2004.
617. H. Takeda, T. Nishimoto, and S. Sagayama. Rhythm and tempo recognition
of musical performance from a probabilistic approach. In International Con-
ference on Music Information Retrieval, Barcelona, Spain, October 2004.
618. D. Talkin. A robust algorithm for pitch tracking. In Kleijn and Paliwal [355],
pp. 495-517.
619. A.S. Tanguiane. Artificial Perception and Music Recognition. Springer, Berlin
Heidelberg, 1993.
620. T. Tarvainen. Automatic Drum. Track Transcription from Polyphonic Mu-
sic. Master's thesis. Department of Computer Science, University of Helsinki,
Helsinki, Finland, May 2004.
621. D. Temperley. The Cognition of Basic Musical Structures. MIT Press, Cam-
bridge, Massachusetts, 2001.
622. D. Temperley and D. Sleator. Modeling meter and harmony: A preference-rule
approach. Computer Music Journal, 23(1): 10-27, Spring 1999.
623. M. Tervaniemi and K. Hugdahl. Lateralization of auditory-cortex functions.
Brain Research Reviews, 43(3):231-46, 2003.
624. D. Thompson, editor. Concise Oxford English Dictionary. Clarendon Press, 9
edition, 1995.
625. H.D. Thornburg, R.J. Leistikow, and J. Berger. Melody retrieval and musi-
cal onset detection from the STFT. IEEE Transactions on Speech and Audio
Processing, 2005. (submitted)
626. P. Toiviainen. An interactive MIDI accompanist. Computer Music Journal,
22(4):63-75, Winter 1998.
627. T. Tolonen and M. Karjalainen. A computationally efficient multipitch analy-
sis model. IEEE Transactions on Speech and Audio Processing, 8(6):708-716,
2000.
628. T. Tolonen, V. Valimaki, and M. Karjalainen. Evaluation of modern sound
synthesis methods. Technical Report 48, Helsinki University of Technology,
March 1998.
629. I. Tsochantaridis, T. Hofmann, T. Joachims, and Y. Altun. Support vector
machine learning for interdependent and structured output spaces. In Interna-
tional Conference on Machine Learning, Banff, Canada, 2004.
630. G. Tzanetakis. Song-specific bootstrapping of singing voice structure. In IEEE
International Conference on Multimedia and Expo, Sorrento (Naples), Italy,
2004.
 References 425

631. G. Tzanetakis and P. Cook. Musical genre classification of audio signals. IEEE
Transactions on Speech and Audio Processing^ 10(5):293-302, July 2002.
632. G. Tzanetakis, G. Essl, and P. Cook. Automatic musical genre classification of
audio signals. In 2nd Annual International Symposium on Music Information
Retrieval, 2001.
633. G. Tzanetakis, G. Essl, and P. Cook. Human perception and computer extrac-
tion of beat strength. In International Conference on Digital Audio Effects,
pp. 257-261, 2002.
634. C. Uhle, C. Dittmar, and T. Sporer. Extraction of drum tracks from polyphonic
music using independent subspace analysis. In the 4th International Symposium
on Independent Component Analysis and Blind Signal Separation, Nara, Japan,
2003.
635. C. Uhle and J. Herre. Estimation of tempo, micro time and time signature
from percussive music. In International Conference on Digital Audio Effects,
London, UK, 2003.
636. M. Unoki and M. Akagi. A method of signal extraction from noisy signal based
on auditory scene analysis. Speech Communication, 27(3):261-279, 1999.
637. L. Van Immerseel and J.P. Martens. Pitch and voiced/unvoiced determina-
tion with an auditory model. Journal of the Acoustical Society of America,
91(6):3511-3526, June 1992.
638. S.V. Vaseghi. Advanced Digital Signal Processing and Noise Reduction. John
Wiley & Sons, 1996.
639. S.V. Vaseghi. Advanced Digital Signal Processing and Noise Reduction,
pp. 270-290. Wiley, 2nd edition, July 2000.
640. R. Ventura-Miravet, F. Murtagh, and J. Ming. Pattern recognition of musi-
cal instruments using hidden Markov models. In Stockholm Music Acoustics
Conference, Stockholm, Sweden, 2003.
641. T. Verma and T. Meng. Extending spectral modeUng synthesis with transient
modeling synthesis. Computer Music Journal, 24(2):47-59, 2000.
642. M. Vetterli and J. Kovacevic. Wavelets and Subband Coding. Prentice Hall,
Englewood Cliffs, NJ, USA, 1995.
643. T. Viitaniemi, A. Klapuri, and A. Eronen. A probabilistic model for the tran-
scription of single-voice melodies. In 2003 Finnish Signal Processing Sympo-
sium, pp. 59-63, Tampere, Finland, May 2003.
644. E. Vincent. Modeles d'Instruments pour la Separation de Sources et la Tran-
scription d'Enregistrements Musicaux. PhD thesis, IRCAM and Universite
Pierre et Marie Curie, Paris, 2004.
645. E. Vincent. Musical source separation using time-frequency source priors. IEEE
Transactions on Acoustics, Speech, and Signal Processing, 2005. special issue
on Statistical and Perceptual Audio Processing, to appear.
646. E. Vincent and M.D. Plumbley. A prototype system for object coding of musical
audio. In IEEE Workshop on Applications of Signal Processing to Audio and
Acoustics, New Paltz, USA, October 2005.
647. E. Vincent and X. Rodet. Instrument identification in solo and ensemble music
using independent subspace analysis. In International Conference on Music
Information Retrieval, Barcelona, Spain, 2004.
648. E. Vincent and X. Rodet. Music transcription with ISA and HMM. In the
5th International Symposium on Independent Component Analysis and Blind
Signal Separation, 2004.
426 References

649. T. Virtanen. Audio Signal Modeling with Sinusoids Plus Noise. Technical re-
port, Tampere University of Technology, Department of Information Technol-
ogy, 2000. Master's thesis.
650. T. Virtanen. Sound source separation using sparse coding with temporal conti-
nuity objective. In International Computer Music Conference, Singapore, 2003.
651. T. Virtanen. Separation of sound sources by convolutive sparse coding. In
ISCA Tutorial and Research Workshop on Statistical and Perceptual Audio
Processing, Jeju, Korea, 2004.
652. T. Virtanen and A. Klapuri. Separation of harmonic sound sources using sinu-
soidal modeling. In IEEE International Conference on Acoustics, Speech, and
Signal Processing, Volume 2, pp. 765-768, Istanbul, Turkey, 2000.
653. T. Virtanen and A. Klapuri. Separation of harmonic sounds using hnear models
for the overtone series. In IEEE International Conference on Acoustics, Speech,
and Signal Processing, Orlando, USA, 2002.
654. A. Viterbi. Error bounds for convolutional codes and an asymptotically
optimum decoding algorithm. IEEE Transactions on Information Theory,
13(2):260-269, April 1967.
655. G.H. Wakefield. Mathematical representation of joint time-chroma distribu-
tions. In SPIE Conference on Advanced Signal Processing Algorithms, Archi-
tectures, and Implementations, pp. 637-645, 1999.
656. P.J. Walmsley. Signal Separation of Musical Instruments: Simulation-Based
Methods for Musical Signal Decomposition and Transcription. PhD thesis. De-
partment of Engineering, University of Cambridge, September 2000.
657. P.J. Walmsley, S.J. Godsill, and P.J.W. Rayner. Multidimensional optimisa-
tion of harmonic signals. In European Signal Processing Conference, Island of
Rhodes, Greece, September 1998.
658. P.J. Walmsley, S.J. Godsill, and P.J.W. Rayner. Polyphonic pitch tracking us-
ing joint Bayesian estimation of multiple frame parameters. In IEEE Workshop
on Applications of Signal Processing to Audio and Acoustics, New Paltz, USA,
October 1999.
659. C. Wang, R. Lyu, and Y. Chiang. A robust singing melody tracker using adap-
tive round semitones (ARS). In 3rd International Symposium on Image and
Signal Processing and Analysis, pp. 549-554, Rome, Italy, 2003.
660. Y. Wang. A beat-pattern based error concealment scheme for music delivery
with burst packet loss. In IEEE International Conference on Multimedia and
Expo, 2001.
661. Y. Wang and M. Vilermo. A compressed domain beat detector using mp3 audio
bitstreams. In ACM International Multimedia Conference, Ottawa, Canada,
2001.
662. R.M. Warren. Perceptual restoration of missing speech sounds. Science,
167:392-393, 1970.
663. M. Weintraub. A computational model for separating two simultaneous talkers.
In IEEE International Conference on Acoustics, Speech, and Signal Processing,
pp. 81-84, Tokyo, Japan, 1986.
664. J. Wellhausen and H. Crysandt. Temporal audio segmentation using MPEG-7
descriptors. In SPIE Storage and Retrieval for Media Databases 2003, Vol-
ume 5021, pp. 380-387, 2003.
665. D. Wessel. Timbre space as a musical control structure. Computer Music Jour-
nal, 3:45-52, 1979.
 References 427

666. H. White. Connectionist nonparametric regression: Multilayer feedforward net-

works can learn arbitrary mappings. Neural Networks, 3:535-549, 1990.
667. M.V. Wickerhauser. Adapted Wavelet Analysis from Theory to Software. AK
Peters, Boston, MA, USA, 1994.
668. A. Wieczorkowska. Classification of musical instrument sounds using decision
trees. In 8th International Symposium on Sound Engineering and Mastering,
pp. 225-230, 1999.
669. A. Wieczorkowska. Rough sets as a tool for audio signal classification. In
Z.W. Ras and A. Skowron, editors, Foundations of Intelligent Systems: Pro-
ceedings of the 11th International Symposium on Foundations of Intelligent
Systems, pp. 367-375. Springer, 1999.
670. A. Wieczorkowska, J. Wroblewski, P. Synak, and D. Slezak. Application of
temporal descriptors to musical instrument sound recognition. Journal of In-
telligent Information Systems, 21:71-93, 2003.
671. A. Wieczorkowska and J.M. Zytkow. Analysis of feature dependencies in sound
description. Journal of Intelligent Information Systems, 20:285-302, 2003.
672. W. Wiegerinck and D. Barber. Mean field theory based on belief networks for
approximate inference. In 8th International Conference on Artificial Neural
Networks, pp. 499-504, Skovde, Sweden, 1998.
673. I.H. Witten and T.C. Bell. The zero-frequency problem: Estimating the prob-
abilities of novel events in adaptive text compression. IEEE Transactions on
Information Theory, 37(4): 1085-1094, 1991.
674. I.H. Witten and E. Frank. Data Mining: Practical Machine Learning Tools and
Techniques. Amsterdam: Elsevier-Morgan Kaufmann, second edition, 2005.
675. E. Wold, T. Blum, D. Keislar, and J. Wheaton. Classification, search and
retrieval of audio. In B. Furth, editor. Handbook of Multimedia Computing,
pp. 207-226. CRC Press, 1999.
676. P.J. Wolfe, S.J. Godsill, and W.J. Ng. Bayesian variable selection and regular-
isation for time-frequency surface estimation. Journal of the Royal Statistical
Society: Series B, 66(3):575-589, August 2004.
677. M. Wu, D. Wang, and G.J. Brown. A multipitch tracking algorithm for noisy
speech. IEEE Transactions on Speech and Audio Processing, 11(3):229-241,
2003. URL: www.cse.ohio-state.edu/ dwang/pnl/ software.html.
678. H. Yamada, M. Goto, H. Saruwatari, and K. Shikano. Multi-timbre chord clas-
sification for musical audio signals. In 2002 Autumn Meeting of the Acoustical
Society of Japan, pp. 641-642, Akita, Japan, September 2002. (in Japanese)
679. H. Yamada, M. Goto, H. Saruwatari, and K. Shikano. Multi-timbre chord clas-
sification method for musical audio signals: Application to musical pieces. In
the 2003 Spring Meeting of the Acoustical Society of Japan, pp. 835-836, Tokyo,
Japan, March 2003. (in Japanese)
680. G.K. Yates. Cochlear structure and function. In Moore [474], pp. 41-74.
681. C. Yeh and A. Robel. A new score function for joint evaluation of multiple
FO hypotheses. In International Conference on Digital Audio Effects, Naples,
Italy, October 2004.
682. K. Yoshii, M. Goto, and H.G. Okuno. Automatic drum sound description for
real-world music using template adaptation and matching methods. In Interna-
tional Conference on Music Information Retrieval, Barcelona, Spain, October
2004.
428 References

683. K. Yoshii, M. Goto, and H.G. Okuno. Drum sound identification for polyphonic
music using template adaptation and matching methods. In ISC A Tutorial
and Research Workshop on Statistical and Perceptual Audio Processing, Jeju,
Korea, October 2004.
684. T. Yoshioka, T. Kitahara, K. Komatani, T. Ogata, and H.G. Okuno. Automatic
chord transcription with concurrent recognition of chord symbols and bound-
aries. In International Conference on Music Information Retrieval, pp. 100-
105, Barcelona, Spain, October 2004.
685. S.J. Young, N.H. Russell, and J.H.S. Thornton. Token passing: A simple con-
ceptual model for connected speech recognition systems. Technical report, De-
partment of Engineering, University of Cambridge, July 1989.
686. B. Zadrozny and C. Elkan. Obtaining calibrated probability estimates from de-
cision trees and naive Bayesian classifiers. In International Conference on Ma-
chine Learning, pp. 202-209, Williamstown, Massachusetts, USA, June 2001.
687. R.J. Zatorre, P. Belin, and V.B. Penhune. Structure and function of auditory
cortex: Music and speech. TRENDS in Cognitive Sciences, 6(l):37-46, 2002.
688. T. Zhang. Instrument classification in polyphonic music based on timbre analy-
sis. In SPIE, Internet Multimedia Management Systems II, pp. 136-147, 2001.
689. Y. Zhu and M.S. Kankanhalli. Robust and efficient pitch tracking for query-
by-humming. In 2003 Joint Conference of the Fourth International Conference
on Information, Communications and Signal Processing, 2003 and the Fourth
Pacific Rim Conference on Multimedia, Volume 3, pp. 1586-1590, Singapore,
December 2003.
690. M. Zibulski and Y. Zeevi. Analysis of multiwindow Gabor-type schemes by
frame methods. Applied and Computational Harmonic Analysis, 4(2): 188-212,
1997.
691. A. Zils. Extraction de Descripteurs Musicaux: Une Approche Volutionniste.
PhD thesis, Sony CSL Paris and Laboratoire d'Informatique de I'universit Paris
6, 2004.
692. A. Zils and F. Pachet. Automatic extraction of music descriptors from acoustic
signals using eds. In Audio Engineering Society 116th Convention, Berlin, Ger-
many, 2004.
693. A. Zils, F. Pachet, O. Delerue, and F. Gouyon. Automatic extraction of drum
tracks from polyphonic music signals. In International Conference on Web
Delivering of Music, Darmstadt, Germany, December 2002.
694. E. Zwicker and H. Fasti. Psychoacoustics: Facts and Models. Springer, 1999.
Index

Accent, 107, 112, 374 frequency response, 237

Acoustic features, see Feature extraction impulse response, 237
Acoustic model, see Source model Auditory model, 6, 14, 234, 372
Adaptive modelling, see Source model of Cooke, 246
adaptation of Ellis, 245
Adaptive oscillator, see Oscillator of Meddis and Hewitt, 241
Additive synthesis, see Sinusoidal model toolboxes, 238
Aerophone, 167, 231 Auditory nerve, 234, 235, 237, 238, 242
Amplitude envelope, 176 Auditory scene analysis, 7, 246, 247,
Amplitude evolution model, 208 299, 306
damped amplitudes, 208, 218 Auditory system, 65, 234, 300
exponential decay model, 207, 219 auditory cortex, 234
Amplitude modulation central auditory processing, 235, 241,
as clue for component fusion, 306 242, 265
tremolo, see Tremolo cochlea, see Cochlea
Amplitude prior, 212 peripheral hearing, 14, 234
Analysis by synthesis, 140 Autocorrelation, 241, 243, 245
Analysis frame, 23, 66, 68 autocorrelation function, 233
Analytic mode of listening, 6 calculation via FFT, 256
Annotation of music, 17, 355 enhanced, 288
Arrhythmic, 105 generalized ACF, 252, 256
Articulation, 9, 193 of a frequency-domain signal, 233
in singing, 364 summary ACF, see Summary
Artificial neural network, see Neural autocorrelation
network use for beat tracking, 112
Attack time, 177 use for FO estimation, 232, 369
Attack-point, 134 Automatic music transcription, see
Audio coding, 3, 70, 88, 91 Transcription
Auditory cortex, 12 Automatic threshold selection, 350
left/right asymmetry, 12 Autoregressive model, 39, 175, 211
Auditory filters, 235 Autoregressive noise, 211
bandwidth, 236 Average magnitude difference function,
centre frequency, 236 251
430 Index

Bandwise energy features, 135 Channel selection

Bar line, 10, 106 in auditory model, 244, 249, 251
Bar line estimation, 124, 159, 341 Choral music, 109, 125
Basilar membrane, 234, 235 Chord, 9, 13
BcLsis function, 269 chord change detection, 117, 328, 341
amplitude, see Gain function notation, 3
frequency domain, 143, 146, 149, 273 recognition, 319
time domain, 271 root detection, 246, 254, 263
Bass line, 13, 321, 329 Chordophone, 167, 231
Bass line transcription, 330 Chorus measure, 345, 353
Bassoon, 167 Chorus sections, 329
Bayes's rule, 38, 40 estimation, 342
Bayesian Chroma and height, 346
classification, 62 Chroma vector, 345, 347
estimation, 40 Clarinet, 96, 166, 167, 176, 177, 206,
model, 203 207, 319
network, 313 Classical music, 7, 112, 125, 128, 340,
Beat, 10, 105, 134, 329, 341 355
Beat induction Classification, 52, 137, 139, 163, 184
by humans, see Metre perception combining, 193
computational, see Beat tracking ensemble, 184
Beat tracking, 6, 101, 341 flat, 184
Beating, 239, 256 generalization, 169, 189
Bell, 167 genre, 13, 112, 356
Binning, 154 hierarchical, 184, 192
Blackboard architecture, 246, 247, 312 unsupervised, 53, 133, 164
Blind source separation, see Source Clustering, 133, 151, 159, 164, 193, 286,
separation 307, 311, 340, 352, 354
Bottom-up processing, 16, 307, 312, Cochlea, 164, 234
342, 352 cochlear filters, see Auditory filters
Box-Cox power transform, 179 computational models, 234
Brain damage, 11 Cocktail party problem, 300
Brain imaging, see Neuroimaging Comb filter, 115, 243, 257
Brass instrument, 167 Combining classifiers, 193
Bugle, 167 Common musical notation, 3, 4
expression symbols, 389
Cancellation filter, 196, 250 Complete family, 72
Canonical discriminant analysis, 180 Complete transcription, 3
Celesta, 167 Compression
Cello, 167, 209, 340 in auditory model, 238, 239, 248
Cent, 332 i/th-law compression, 255
Central auditory processing, see Computational auditory scene analysis,
Auditory system see Auditory scene analysis
Central limit theorem, 274 Concurrent vowel identification, 249
Cepstral coefficients, 25, 26 Conditional probability density
MFCC, see Mel-frequency cepstral function, 31
coefficients Conjugate prior, 47
Cepstrum, 25, 253 Contextual information, 198
Cepstrum pitch detection, 253, 264, 371 in auditory scene analysis, 313
 Index 431

in instrument classification, 166 Discrete Fourier transform, see Fourier

in singing, 385 transform, 273
short-term acoustic context, 154 Discrete wavelet transform, see Wavelet
tonal context, 320, 385 transform
Continuous-length evaluation measure, Discriminant analysis, 61, 185
125 Dissonance, 9, 287
Contrabass, 167, 198 Distribution, see Probability density
Convolution, 22, 25, 289 function
time-frequency, 25 Divergence, 282
Correlation-based feature selection, 182 Double bass, see Contrabass
Correlogram, 227, 242, 245, 339 Drums, see Percussion
Covariance matrix, 29, 30 Dual frame, 78
empirical, 53, 61 Duplex theory of pitch perception, 241
Cramer-Rao bound, 34 Duration
Creative Commons initiative, 170 of a note, 362
Critical band, 236, 274 perceived duration, 8
Critical-band scale, 236 Dyadic grid, 74
Cross-validation, 170 Dynamic model
finite, discrete state space, 63
Damped amplitudes, see Amplitude Dynamic range
evolution model of hearing, 238, 273
Damping factor, 208 of inner hair cells, 238
Dance music, 102, 125, 160
ballroom dance, 117 Ear training, 12
Data representation, see Mid-level data Eardrum, 234
representation Eigenvalue, 54, 180
Data-driven processing, 16 Eigenvector, 54
Databases, 16 Electric guitar, see Guitar
of musical instrument sounds, 169, Elementary waveform, 70
260, 355 EM algorithm, see Expectation
RWC Music Database, 355 maximization
DCT-IV transform, 75 Empirical average, 29
decibel, 8, 294 Empirical probability density function,
Decision tree, 138, 186 82
Decorrelation, 275 Empirical risk, 56
nonlinear, 275 English horn, 167
Delta features, 175 Enhanced summary ACF, see Summary
Delta-MFCC, see Mel-frequency autocorrelation function
cepstral coefficients Ensemble of classifiers, 184
Description of music signal, 328 Equal-tempered scale, 9, 224
Deterministic approach to signal Equivalent rectangular bandwidth, 236
modelling, 67 Estimation,
Diatonic scale, 271, 293 bias, 33, 41
Dictation, see Human transcription covariance, 33
Dictionary of waveforms, see Waveform unbiased estimator, 34
dictionary see also Fundamental frequency
Dilation, 71 estimation, 21
Directed acyclic graph, 120, 313 Estimation theory, 33
Discrete cosine transform, 75 Euclidean distance, 282
432 Index

Evaluation methodology, 170 multiple FO estimation, 203, 229, 248,

Expectation, 29 287
Expectation maximization, 35, 221, 336 of melody and bass fines, 330
Exponential decay, see Amplitude single FO estimation, 232, 369
evolution model typical errors, 233, 253
Expression, 193 Fundamental frequency pdf, 332
estimation and encoding, 389 Fundamental frequency, term definition.

FastICA algorithm, 275

Feature extraction, 66, 135, 171, 247, g-prior, 212
270, 287, 307, 344, 345, 368 Gabor frame, 78
delta, 175 Gabor function, 78
sliding window, 155 Gabor representation, 24, 209
Feature scale transformation, 178 Gabor waveforms, 71
Feature selection, 137, 181 Gain function, 143, 269
Feature transformation, 191 Gamma prior, 211
Fisher ratio, 61 Gammatone filter, 237, 241, 247, 255,
Flat classifier, 184 306
Flute, 205, 206, 209, 319, 323, 340 efficient implementation, 238
Folk music, 125, 323 Gaussian distribution, 28
Formant, 253, 294, 365 Gaussianity of variable, 274
Fourier transform, 21, 22 generalized, 286
continuous, 22 truncated, 213
discrete, 22 Gaussian mixture model, 35, 37, 139,
inverse, 22 191, 385, 386
Fractional-delay filter, 250 Gaussian noise, 279, 282
Frame, see Analysis frame Generalization in classification, 169, 189
Frame (family of vectors), 77 Generalized autocorrelation, see
dual, 78 Autocorrelation
frame inversion, 78 Generative model, 62, 112, 122, 204,
frame representation, 78 277, 324
Gabor frame, 78 Generative signal, speech and music, 16
hybrid system, 81 Genre classification, 13, 112, 356
multiple Gabor frames, 79 Gestalt psychology, 300
overcomplete system, 78 Glissando, 389
quilted frame, 81 in singing, 367
French horn, 166, 167 Glockenspiel, 232
Frequency evolution model, 209 Greedy algorithm, 84, 89, 225, 354
Frequency grid, 219, 224 Grid, temporal, see Temporal grid
Frequency modulation Grouping, see Rhythmic grouping
as clue for component fusion, 306 Guitar, 79, 166, 167, 177, 227
vibrato, see Vibrato electric, 198, 200, 209
Frequency proposal distribution, 216 onset detection, 108
Frequency warping, 254 sound separation, 270, 290
Frobenius norm, 280 transcription, 301
Front end, 244
Full-wave ^'th-law compression, 255 Hair cell, see Inner hair cell. Outer hair
Fundamental frequency estimation cell
basic principles, 232 Half-wave rectification, 239, 256, 258
 Index 433

Harmonic matching pursuit, 92 inertia ratio maximization using feature

Harmonic partial, 107, 231, 291 space projection, 183
Harmonic pattern matching, 233 Information gain, 182
Harmonic selection, 257 Information retrieval, see Music
Harmonic sound, 231 information retrieval
Harmonic trajectory, 225 Inharmonicity, 175, 206, 222
Harmonicity, as clue for component in string instruments, 231
fusion, 306 inharmonicity factor, 232
Harmonics and Individual Lines plus model for piano, 206, 222, 232
Noise, 69 Inner ear, 234
Harmony, 9 Inner hair cell, 234, 235
Harp, 167 models, 238, 241, 246, 247, 255
Harpsichord, 167 Inner lines in music, 6, 13
transcription, 282 Inner product, 71
Heisenberg-Gabor inequality, 24 Inner product space, 71
Hidden Markov model, 63, 121, 122, Instance-based classification, 138, 185
139, 191, 286, 380 /c-NN, see /c-nearest neighbours
Hierarchical beat structure, see Metre Instantaneous frequency, 246, 332
Hierarchical classifier, 184, 192 Instrument
Higher-order statistics, 275, 287 sample databases, see Databases
History of automatic music transcrip- sounds, see Musical sounds
tion, 6, 301 Instrument classification, 7, 65, 163
Human auditory system, see Auditory in humans, 164
system percussion, see Percussion sound
Human transcription, 5, 12 recognition
Human-computer interaction, 341 Instrument families, 167
Integration of information, 16, 154, 313,
Human-computer interaction, 5, 361
Hybrid representation, 67 383
Hybrid system, see Frame (family of Inter-onset interval, 8, 115
Intermediate data representation, see
vectors)
Mid-level data representation
Hyperparameter, 47, 62
Internal model, 15
Interval, 12, 321, 340, 341, 362
i.i.d., 32 Invariant feature extraction, 276, 277
Idiophone, 131, 145, 167 Inverse Fourier transform, see Fourier
Ill-posed problem, 38 transform
Importance probability density Inverse gamma distribution, 47
function, 49 Iterative FO estimation and cancellation,
Importance sampling, 49, 223 196, 250, 254, 259, 263
Independence, see Statistical indepen-
dence JADE algorithm, 275
Independent component analysis, 109, Jazz music, 105, 125, 127, 321, 355
143, 274, 305 Joint estimation of multiple FOs, 214,
multidimensional, 276, 277 251, 252, 263
non-negative, 278 Joint probability density function, 31
spatiotemporal, 277 Junction tree algorithm, 316
Independent subspace analysis, 144,
276, 285 Arnearest neighbours, 60, 185, 191
sub-band, 145 Kahmba, 167
434 Index

Kalman filter, 51, 52, 117, 121, 221-223, Markov chain Monte Carlo, 43, 117,
226, 324 121, 215
Kernel, 57 Markov tree, 90
Gaussian, 57 Masking, 236
positive definite, 57 Matching pursuit, 84
Kullback-Leibler divergence, 282 Matrix diagonalization, 54
symmetric, 287 Matrix factorization, see Non-negative
Kullback-Leibler information, 335 matrix factorization
Kurtosis, 275 Maximum a posteriori estimation, 40,
spectal, 136 121, 214, 279, 331
Maximum likelihood estimation, 33, 35
Language model, 15 Mbira, 167
Laplace approximation, 225 McNemar's test, 171
Laplacian distribution, 280 Mean, 29, 30
Latent variable, 35 empirical, 54, 61
Law of large numbers, 30 Mean square error of an estimator, 34
Lazy learning, 185 Measure
Least-squares, 285 musical measure, see Bar line
Leave-one-out cross-validation, 170 musical measure estimation, see Bar
Lebesgue measure, 28 line estimation
Legato, 9 Mechanical-to-neural transduction, 238
in singing, 367 Mel frequency cepstral coefficients, 26
Level adaptation in auditory model, Mel frequency scale, 26
239, 246 Mel-frequency cepstral coefficients, 63,
Level compression, see Compression 135, 174
Likelihood, 210, 219, 332 delta-MFCC, 175
Likelihood function, 32 mel-frequency cepstral coefficients, 270
degenerate, 33, 38 Mel-frequency scale, 173
penalized, 38, 41, 57 Mel-scale filterbank, 26
Linear discriminant analysis, 186, 191 mel-scale filterbank, 173
Linear interpolation, 253 Melodic phrase, see Phrase
Linear prediction, 39 Melody, 9, 12, 13, 329
Linear programming, 286 perceptual coherence, 15
Local cosine basis, 73 segregation of melodic lines, 247
Localized source model, 142 transcription, see Predominant FO
Locally harmonic sources, 69 estimation
Log-Gaussian distribution, 224 Membranophone, 131, 145, 147, 167
Loss function, 55 Memory for music, 11, 13
Loudness, 8, 172 Message passing, 221
of instrument sounds, 319 Metadata, see Annotation
of melody vs. accompaniment, 340 Metre, 10, 105, 312, 329, 341
of singing, 366 Metre analysis, 101, 134, 341, 388
Metre perception, 10, 11, 102
Mallet percussion instrument, 232 MFCC, see Mel frequency cepstral
Marginal MMSE, 214 coefficients
Marginal probability density function, Mid-level data representation, 12, 13,
31 65, 244, 248, 251, 256, 264
Marimba, 167, 232 desirable qualities, 14
Markov chain, see A^-gram model hybrid, see Hybrid representation
 Index 435

Middle ear, 234 Music-synchronized computer graphics,

MIDI, 3, 4, 9, 101, 102 356
MIDI note number, 362 Musical context, see Contextual
Minimum mean square error estimation, information
41, 214 Musical instrument, see Instrument
Missing feature theory, 197 Musical instrument classification, see
Mixing matrix, 270 Instrument classification
Mixture-of-experts approach, 193 Musical key, 9, 385
Model adaptation, source, see Source change, 344
model adaptation estimation, 386
Model selection, 39 Musical metre, see Metre
Model, signal, see Signal model Musical scale, see Scale
Modified discrete cosine transform, 75 Musical sounds
MDCT basis, 73 percussive, 107, 131
Modularity, 11 pitched, 107, 167, 231
Modulation, 71 Musicological modeUing, 15, 153-155
amplitude, see Amplitude modulation melodic continuity, 331
frequency, see Frequency modulation musical key, 385
Modulation (musical key change), 344 of periodic patterns, 157
Molecular matching pursuit, 94 of rhythmic patterns, 159
Moment order r, 30 short-term context modelling, 154
Monophonic signal, 5 with N-grains, see AT-gram model
Monte Carlo, 41 Mutual information, 274
MPEG-7, 136, 167, 172, 270, 354
Multi-class classification, 55
Multi-layer perceptron, 187, 191 AT-fold cross-validation, 170
Multidimensional scaling, 168, 179 N-gram model, 155
Multiple FO (non)stationary model, see for chord sequences, 319
Signal model of melody, 321, 387
Multiple FO estimation, see Fundamen- of percussion sequences, 155, 156
tal frequency estimation periodic iV-gram, 157, 158
Multiplicative update rule, 283 Neural firing probability, 237, 238
Multiresolution analysis, 74 Neural impulse, 234, 238, 242
Multiwavelet, 77 Neural network, 187, 191
Music cognition, 11 MLP, 187, 191
impaired cognition, 11 time-delay neural network, 247
Music information retrieval, 5, 102, 170, Neural spike, see Neural impulse
327, 356, 363 Neuroimaging, 11
Music listening station, 357 Neurophysiology of music cognition, 11
Music map, 357 Noise, 32
Music perception, 5, 103, 327 autoregressive, 211
Music scene analysis, 299 Gaussian, 279, 282
Music scene description, 327, 328 Poisson, 282
Music structure, see Structure white, 32, 39
Music structure analysis, see Structure Noise robustness in FO estimation, 251,
analysis 253, 264
Music thumbnail, 342 Noisy sum-of-sines model, see Signal
Music transcription, see Transcription model
Music-playback interface, 357 Non-negative matrix deconvolution, 293
436 Index

Non-negative matrix factorization, 148, Particle filter, 50, 117, 121

282 Rao-BlackwelUzed, 222
Norm, 71 Peak selection in autocorrelation
Normal distribution, see Gaussian function, 251
distribution Penalized likelihood, see Likelihood
Notation, see Common musical notation function
Note, 3, 362 Perception
MIDI note number, 362 of metre, see Metre perception
Note birth move, 216 of music, see Music perception
Note death move, 216 of pitch, see Pitch perception
Note labelling in singing transcription, Perceptual attributes of sounds, 8
362, 378 Perceptual categorization, 9
Note model, see Source model Perceptual sound vs. physical sound,
Note segmentation, see Temporal 302
segmentation Percussion notation, 3
Note update move, 216 Percussion sound recognition, 133, 137,
Nuisance parameters, 220 174, 184
Nyquist frequency, 23 clustering and labelling, 159
Percussion transcription, 6, 7, 131, 329,
Oboe, 167, 209 342
Observation density, see Likelihood pattern recognition-based, 133
Octave equivalence, 9 separation-based, 142
Odd-to-even ratio, 175 Periodicity
Offset (a)synchrony, 306 in the frequency domain, 233
Onset, 101, 260, 262 in the time domain, 232
and beat, 106 Peripheral hearing, see Auditory system
of percussive sounds, 108 Phase generation, 288
of pitched sounds, 108 Phenomenal accent, see Accent
Onset (a)synchrony, 306 Phonation, 364
Onset detection, 102, 107, 134, 195, 287 frequency, 365
Organ of Corti, 234 types, 365
Orthogonal matching pursuit, 85 Phoneme, 16
Orthonormal basis, 72 Phrase, 10
Oscillator, 113 Phrasing, 193
adaptive, 113, 247 Physical sound vs. perceptual sound,
comb filter, 115, 244, 257 302
oscillator net, 247 Piano, 24, 167, 203, 206, 207, 209, 319,
Outer ear, 234 323
Outer hair cell, 234 identification in music, 198
Overcomplete system, see Frame inharmonicity, see Inharmonicity
(family of vectors) keyboard, 9
Overlap-add, 288 onset detection, 108
Overtone partial, see Harmonic partial transcription, 245, 247
Piano roll, 4, 220
Parametric model of signal, see Signal Pitch, 8, 107
model of noise signals, 229
Parseval formula, 73 perception, see Pitch perception
Partial de-tuning, 206, 306 tonal encoding, see Tonal encoding
Partial transcription, 3 zoo of pitch effects, 229
 Index 437

Pitch class, 9, 345, 385 Pulse, see Beat; Metre

Pitch label of a note, 362 Pulse Code Modulation (PCM), 65
Pitch perception, 229, 242
Pitch perception model, 234, 241 Quantization, 9
autocorrelation model, 241, 242 of FO values, 213, 362
duplex theory, 241 of onset times, 101, 106, 118, 389
shortcomings, 248, 262 of signal sample values, 65
unitary model, 242, 252 Query by humming, 356, 361
Pitched musical sounds, 107, 167, 231 Quilted frame, see Frame (family of
Pizzicato, 192 vectors)
Poisson noise, 282
Poisson prior, 214 Random variable, 28
Polyphonic signal, 5 expectation, see Expectation
Popular music, 10, 125, 128, 131, 137, i.i.d., see i.i.d.
151, 329, 330, 342, 344, 345, 353, independent, see Statistical indepen-
355 dence
Predominant FO estimation, 198, 260, Random walk, 208
330 Rayleigh quotient, 61
PreFEst method, 330 Recognition
Principal component analysis, 54, 137, of percussion sounds, see Percussion
139, 144, 151, 152, 155, 179, 181, sound recognition
275 of pitched sounds, see Instrument
Prior, 40 classification
conjugate, 47 Rectification, see Half-wave rectification
sequential, 217, 220, 221, 225 Redundancy, 78
Prior distribution, 40 Reed instrument, 167
Prior selection, 211 RefraiD method, 345
Prior subspace analysis, 146, 150 Register (in singing), 365
input generated priors, 151 Regular grid, 75
non-negative, 148 Regularized risk, 57, 58
Probabilistic model, 31 Repeated sections, see Structure
Probability density function, 28 Reproducing kernel Hilbert space, 57
conditional, 31 Resolvability, 258
empirical, 82 Rest (in music), 362
Gaussian, see Gaussian distribution Rhythm, 10, 105
importance, 49 analysis, see Metre analysis
inverse gamma, 47 Rhythmic grouping, 10, 105
joint, 31 Rhythmic pattern modelhng, 159
Laplacian, 280 Risk, 55
marginal, 31 empirical, 56
normal, see Gaussian distribution regularized, 57, 58
prior, 40 Rock music, 105, 112, 125, 131, 151
proposal, 44 Root-mean-square level, 172
uniform, 28 Rotation matrix, 218
Projected steepest descent algorithm, Royalty-free music, 355
281 RWC Music Database, 355
Proposal distribution, 44
Pseudoinverse, 147 Salience
Psychoacoustics, 8, 173, 235, 308 of an onset, 107
438 Index

of FO candidate, 257 Singular value decomposition, 277

Sampling, 65 Sinusoidal model, 14, 68, 225, 246
Saxophone, 167, 198, 207, 256, 259 other variants, see Signal model
Scale, 10, 385 single sinusoid, 31
equal-tempered, 9, 224 Sinusoids -\- transients -h noise model,
Scale tone, 10, 385 70
Scale transformation, 178 Sliding window method, 155
Scaling function, 74 SmartMusicKIOSK, 357
Scene analysis, 299, see also Auditory Sound production mechanism
scene analysis in musical instruments, 167
Schema-based segregation, 312 in singing, 364
Score, see Common musical notation Sound source separation, see Source
Scoretime, 105 separation
Segmentation, see Temporal segmenta- Source model, 15
tion percussion sounds, 149
Semitone, 291, 292, 332, 362 pitched sounds, 319
Sequential dependency, see N-gram statistical note model, 379
model tone model, 319, 333
Sequential Monte Carlo, 117
Source model adaptation, 140, 142, 197,
Sequential prior, 217, 220, 221, 225
333
Shannon theorem, 23
in time domain, 140
Short-time Fourier transform, 23
in time-frequency domain, 141
Signal model, 66, 204
Source separation, 15, 65, 142, 143, 195,
multiple FO non-stationary, 209
249, 267, 304, 305, 327
noisy sum-of-sines, 204
Sparse coding, 143, 278, 285
parametric, 67, 70
non-negative, 153, 281, 285
single FO non-stationary, 207
Sparse expansion, 82
single FO stationary, 204
Sparse representation, 67
sinusoidal, see Sinusoidal model
Spatial information, 267, 306
sum-of-sines model, 204
Spectral features
weighted mixture of tone models, 332
Signal space, 71 spectral centroid, 136
Significance map, 88 spectral flatness, 173
Significance tree, 89 spectral flux, 174
Similarity matrix, 349 spectral irregularity, 174
Simplex algorithm, 286 spectral kurtosis, 136
Singer's formant, 366 spectral rolloff, 174
Singing spectral shape, 135
acoustic characteristics, 364 spectral skewness, 136
production of singing sounds, 364 spectral spread, 136
singing out of tune, 367 Spectral model synthesis, 69
Singing transcription Spectral organization, see Auditory
applications, 361 scene analysis
expression encoding, 389 Spectral smoothness, 207, 240
problem formulation, 361 Spectral whitening, 239, 248, 252, 255,
segment-and-label approach, 377 263
statistical approach, 379 Spectrogram, 23, 24, 65
Single FO (non)stationary model, see Spectrogram factorization, 143, 268
Signal model Spectrum envelope, 207
 Index 439

Speech of instrument sounds, 177

recognition, 15, 16, 63, 156, 307, 358 of singing notes, 362, 373, 377
speech separation, 249 Threshold of hearing, 273
speech signals, 16 Thumbnail, 342
Squared difference function, 250 Tick, see Tatum
Staccato, 9 Timbre, 8, 231
Statistical independence, 31, 274 acoustic correlates, 168
Statistical significance tests, 170 acoustic features, see Feature
Stiffness of vibrating strings, 232 extraction
Stochastic signal components, 69 FO dependency, 189
Strobed temporal integration, 244 of singing sounds, 365
Structure (of a musical work), 10, 329 perceived similarity, 168
Structure analysis, 10 use for auditory organization, 306
by humans, 13 Timbre space, 168
computational, 342 Time quantization, see Quantization
Structured approximation, 87 Time-frequency molecule, 87
Structured audio coding, 3 Time-frequency atom, 70
Student-Fisher t-test, 170 Time-frequency covariance, 24
Style detection, see Genre classification Time-frequency jigsaw puzzle, 85
Sub-band coding, 74 Time-frequency lattice, 209
Sub-beat structure, 103, 106, 117 Time-frequency representation, 13, 23,
Subglottic pressure, 364 25, 65
Sum-of-sines model, 204 Time-lag triangle, 349
Summary autocorrelation function, 242, Time-persistence, 88
245, 248 Timing deviations, 105, 375
enhanced summary ACF, 253 Timpani, 167
Super-beat structure, 103, 106 Token-passing algorithm, 380
Supervised classification, 55, 164 Tonal encoding of pitch, 10, 11
Support vector machine, 57, 138, 191 Tonal music, 10, 304, 330, 385
output moderating, 154 Tone model, see Source model
Swing, 105 Tonic note, 10
Symbolic representation, 9, 102, 107, Top-down processing, 16, 312, 342, 352
302 Transcription
Synchrony strand, 247 by humans, 5, 12
Synthesis of separated sources, 288 complete vs. partial, 3
designing transcription system, 11
Tactus, 10, 105 state of the art, 7
Tatum, 106, 134 subtopics, 5, 11
analysis, 116, 124, 134, 159 trends and approaches, 6, 301
Taxonomy, 164 Transient, 70
of musical instruments, 167 Transientness index, 89
Template matching, 139 Translation, 71
Tempo, 6, 101, 105 Tremolo, 178, 389
estimation, 103, 118, 354, 374 in singing, 367
variation, 105, 106, 388 Tristimulus, 176
Tempogram, 121 Trombone, 166, 167
Temporal centroid, 136, 178 Trumpet, 167, 176, 198, 209, 245, 254,
Temporal grid, 134 319
Temporal segmentation, 12, 134 Tuba, 167
440 Index

Tuning, 9 Vocal organ, 364

absolute, 362 Voice source, 364
drift of, 368 Voicing, degree of, 373
in note labelling, 378
singing out of tune, 367 Waveform dictionary, 67, 81
Two-way mismatch, 288, 371 Waveform representation, 13, 72
Wavelet, 71
Ukulele, 167 Wavelet basis, 73
Unbiased estimator, see Estimation Wavelet transform, 74
Uniform distribution, 28 Weighted-mixture model, see Signal
Unitary model, see Pitch perception
model
model
Well-tempered scale, see Equal-
Unmixing matrix, 147, 275
tempered scale
Unsupervised
Western music, 3, 5, 9, 10, 231, 304
classification, 53, 133, 164
White noise, 32, 39
clustering, see Clustering
Whitening, see Spectral whitening
learning, 7, 16, 267, 380, 382
Wiener-Khintchine theorem, 252
Validation, 169, 170 Wigner-Ville representation, 25
Vibraphone, 232 Window function, 23, 75, 208
Vibrating bar, 232 Windowing, 23
Vibrato, 178, 389 Woodwind instrument, 167
in singing, 366 Written music, see Common musical
rate and depth estimation, 389 notation
Viola, 167
Violin, 9, 166, 167, 177, 207, 209, 231, Xylophone, 167, 232
254, 259, 319, 323
onset detection, 109 Zero crossing rate, 136, 174, 182
Viterbi algorithm, 63, 122, 225, 380 Zero tree, 87