Ph.D.

thesis
Computationally efficient methods
for polyphonic music transcription

Antonio Pertusa Ibáñez

Supervised by
José Manuel Iñesta Quereda
External reviewers:

Anssi Klapuri (Queen Mary University, London, UK)

Andreas Rauber (Vienna University of Technology, Austria)

Committee members:
Xavier Serra (Universitat Pompeu Fabra, Barcelona, Spain)
Gérard Assayag (IRCAM, Paris, France)
Anssi Klapuri (Queen Mary University, London, UK)
José Oncina (Universidad de Alicante, Spain)
Isabel Barbancho (Universidad de Málaga, Spain)
A Teima
 Acknowledgments

First and foremost, I would like to thank all members of the computer music lab
from the University of Alicante for providing an excellent, inspiring, and pleasant
working atmosphere. Especially, to the head of the group and supervisor of this
work, Prof. José Manuel Iñesta. His encouraging scientific spirit provides an
excellent framework for inspiring the new ideas that make us to continuously
grow and advance. I own this work to his advice, support and help.
Carrying out a PhD is not an easy task without the help of so many people.
First, I would like to thank all the wonderful staff of our GRFIA group, and
in general, all the DLSI department from the University of Alicante. My
research periods at the Audio Research Group from the Tampere University of
Technology, the Music Technology Group from the Universitat Pompeu Fabra,
and the Department of Software Technology and Interactive Systems from the
Vienna University of Technology, also contributed decisively to make this work
possible. I have learned much, as a scientist and as a person, from the wonderful
and nice researchers of all these labs.
I would also thank to the people who directly contributed to this work.
I am grateful to Dr. Francisco Moreno for delaying some of my teaching
responsibilities when this work was in progress, and for supplying the kNN
algorithms code. I learned most of the signal processing techniques needed for
music transcription from Prof. Anssi Klapuri. I’ll always be very grateful for the
great period in Tampere and his kind hosting. He directly contributed to this
dissertation providing the basis for the sinusoidal likeness measure code, and also
the multiple f0 databases that allowed to evaluate and improve the proposed
algorithms. Thanks must also go to one of my undergraduate students, Jasón
Box, which collaborated to this work building the ODB database and migrating
the onset detection code from C++ into D2K.
I wish to express my gratitude to the referees of this dissertation, for kindly
accepting the review process, and to the committee members.
This work would not have been possible without the primary support
provided by the Spanish PROSEMUS project1 and the Consolider Ingenio
2010 MIPRCV research program2 . It has also been funded by the Spanish
CICYT projects TAR3 and TIRIG4 , and partially supported by European
Union-FEDER funds and the Generalitat Valenciana projects GV04B-541 and
GV06/166.
Beyond research, I would like to thank my family and my friends (too many
to list here, you know who you are). Although they don’t exactly know what
1 Code TIN2006-14932-C02
2 Code CSD2007-00018
3 Code TIC2000-1703-CO3-02
4 Code TIC2003-08496-C04

v
I am working on and will never read a boring technical report in English, their
permanent understanding and friendship have actively contributed to keep my
mind alive within this period.
Finally, this dissertation is dedicated to the most important person in my
life, Teima, for her love, support, care and patience during this period.

Antonio Pertusa Ibáñez

February, 2010

vi
 Contents

1 Introduction 1

2 Background 7
2.1 Analysis of audio signals . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Fourier transform . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Time-frequency representations . . . . . . . . . . . . . . . 11
2.1.3 Filters in the frequency domain . . . . . . . . . . . . . . . 15
2.2 Analysis of musical signals . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.2 Timbre . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.3 Taxonomy of musical instruments . . . . . . . . . . . . . . 20
2.2.4 Pitched musical sounds . . . . . . . . . . . . . . . . . . . 21
2.2.5 Unpitched musical sounds . . . . . . . . . . . . . . . . . . 24
2.2.6 Singing sounds . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Music background . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.3.1 Tonal structure . . . . . . . . . . . . . . . . . . . . . . . . 27
2.3.2 Rhythm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.3.3 Modern music notation . . . . . . . . . . . . . . . . . . . 33
2.3.4 Computer music notation . . . . . . . . . . . . . . . . . . 34
2.4 Supervised learning . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.1 Neural networks . . . . . . . . . . . . . . . . . . . . . . . 38
2.4.2 Nearest neighbors . . . . . . . . . . . . . . . . . . . . . . 40

3 Music transcription 43
3.1 Human music transcription . . . . . . . . . . . . . . . . . . . . . 43
3.2 Multiple fundamental frequency estimation . . . . . . . . . . . . 45
3.2.1 Harmonic overlap . . . . . . . . . . . . . . . . . . . . . . . 46
3.2.2 Beating . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.3 Evaluation metrics . . . . . . . . . . . . . . . . . . . . . . 48
3.3 Onset detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.3.1 Evaluation metrics . . . . . . . . . . . . . . . . . . . . . . 53

4 State of the art 55

4.1 Single fundamental frequency estimation . . . . . . . . . . . . . . 55
4.1.1 Time domain methods . . . . . . . . . . . . . . . . . . . . 55
4.1.2 Frequency domain methods . . . . . . . . . . . . . . . . . 56
4.1.3 Perceptual models . . . . . . . . . . . . . . . . . . . . . . 59
4.1.4 Probabilistic models . . . . . . . . . . . . . . . . . . . . . 60
4.2 Multiple fundamental frequency estimation . . . . . . . . . . . . 62
4.2.1 Salience methods . . . . . . . . . . . . . . . . . . . . . . . 62

vii
CONTENTS

4.2.2 Iterative cancellation methods . . . . . . . . . . . . . . . 63

4.2.3 Joint estimation methods . . . . . . . . . . . . . . . . . . 65
4.2.4 Supervised learning methods . . . . . . . . . . . . . . . . 66
4.2.5 Unsupervised learning methods . . . . . . . . . . . . . . . 69
4.2.6 Matching Pursuit methods . . . . . . . . . . . . . . . . . 70
4.2.7 Bayesian models . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.8 Statistical spectral models . . . . . . . . . . . . . . . . . . 72
4.2.9 Blackboard systems . . . . . . . . . . . . . . . . . . . . . 74
4.2.10 Database matching . . . . . . . . . . . . . . . . . . . . . . 75
4.3 Discussion of multiple f0 estimation methods . . . . . . . . . . . 76
4.4 Onset detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.4.1 Signal processing methods . . . . . . . . . . . . . . . . . . 77
4.4.2 Machine learning methods . . . . . . . . . . . . . . . . . . 81
4.5 Discussion of onset detection methods . . . . . . . . . . . . . . . 82

5 Onset detection 83
5.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.1.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.1.2 Onset detection functions . . . . . . . . . . . . . . . . . . 85
5.1.3 Peak detection and thresholding . . . . . . . . . . . . . . 89
5.2 Evaluation with the ODB database . . . . . . . . . . . . . . . . . 89
5.2.1 Results using o[t] . . . . . . . . . . . . . . . . . . . . . . . 89
5.2.2 Results using õ[t] . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 MIREX evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.3.1 Methods submitted to MIREX 2009 . . . . . . . . . . . . 93
5.3.2 MIREX 2009 onset detection results . . . . . . . . . . . . 95
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

6 Multiple pitch estimation using supervised learning 101

6.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.1.1 Construction of the input-output pairs . . . . . . . . . . . 102
6.2 Supervised methods . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.2.1 Time-delay neural networks . . . . . . . . . . . . . . . . . 104
6.2.2 Nearest neighbors . . . . . . . . . . . . . . . . . . . . . . 105
6.3 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.3.1 Generation of the training and test sets . . . . . . . . . . 107
6.4 Results using time-delay neural networks . . . . . . . . . . . . . . 108
6.4.1 Neural network parametrization . . . . . . . . . . . . . . 108
6.4.2 Recognition results . . . . . . . . . . . . . . . . . . . . . . 109
6.4.3 Changing waveshapes for detection . . . . . . . . . . . . . 111
6.5 Results using k nearest neighbors . . . . . . . . . . . . . . . . . . 113
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

viii
 CONTENTS

7 Multiple f0 estimation using signal processing methods 119

7.1 Iterative cancellation method . . . . . . . . . . . . . . . . . . . . 120
7.1.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . 120
7.1.2 Onset detection . . . . . . . . . . . . . . . . . . . . . . . . 123
7.1.3 Candidate selection . . . . . . . . . . . . . . . . . . . . . 123
7.1.4 Iterative cancellation . . . . . . . . . . . . . . . . . . . . . 123
7.1.5 Postprocessing . . . . . . . . . . . . . . . . . . . . . . . . 126
7.2 Joint estimation method I . . . . . . . . . . . . . . . . . . . . . . 126
7.2.1 Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . 127
7.2.2 Candidate selection . . . . . . . . . . . . . . . . . . . . . 128
7.2.3 Generation of combinations of candidates . . . . . . . . . 129
7.2.4 HPS estimation . . . . . . . . . . . . . . . . . . . . . . . . 129
7.2.5 Salience of a combination . . . . . . . . . . . . . . . . . . 131
7.2.6 Postprocessing . . . . . . . . . . . . . . . . . . . . . . . . 133
7.3 Joint estimation method II . . . . . . . . . . . . . . . . . . . . . 133
7.3.1 Temporal smoothing . . . . . . . . . . . . . . . . . . . . . 134
7.3.2 Partial search . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.3.3 Fundamental frequency tracking . . . . . . . . . . . . . . 138
7.3.4 Alternative architectures . . . . . . . . . . . . . . . . . . . 139
7.4 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.4.1 Parametrization . . . . . . . . . . . . . . . . . . . . . . . 141
7.4.2 Results using random mixtures . . . . . . . . . . . . . . . 149
7.4.3 Evaluation and comparison with other methods . . . . . . 157
7.4.4 Overall MIREX comparison . . . . . . . . . . . . . . . . . 162
7.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168

8 Conclusions and future work 171

8.1 Discussion and future lines of work . . . . . . . . . . . . . . . . . 172
8.2 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174

A Resumen 177

Bibliography 191

ix
 List of Figures

1.1 Music transcription example . . . . . . . . . . . . . . . . . . . . . 3

2.1 Complex plane diagram . . . . . . . . . . . . . . . . . . . . . . . 9

2.2 Example spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Example magnitude spectrogram . . . . . . . . . . . . . . . . . . 12
2.4 Wavelet filter bank . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.5 Time-frequency resolution grids . . . . . . . . . . . . . . . . . . . 14
2.6 Example filter bank . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.7 Mel scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.8 Attack, sustain and release . . . . . . . . . . . . . . . . . . . . . 18
2.9 Harmonic spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.10 Piano waveform and spectrogram . . . . . . . . . . . . . . . . . . 22
2.11 Vibraphone waveform and spectrogram . . . . . . . . . . . . . . 25
2.12 Snare waveform and spectrogram . . . . . . . . . . . . . . . . . . 25
2.13 Voice waveform and spectrogram . . . . . . . . . . . . . . . . . . 27
2.14 Western musical notes . . . . . . . . . . . . . . . . . . . . . . . . 28
2.15 Musical keys . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.16 Harmonics and intervals. Fig. from Krumhansl (2004) . . . . . . 30
2.17 Metrical levels and timing. Fig. from Hainsworth (2003) . . . . . 32
2.18 Modern notation example . . . . . . . . . . . . . . . . . . . . . . 35
2.19 Note names, score location, frequencies and MIDI pitches . . . . 36
2.20 MIDI visual representations example . . . . . . . . . . . . . . . . 38
2.21 Multilayer perceptron architecture . . . . . . . . . . . . . . . . . 39
2.22 TDNN architecture. Fig. from Duda et al. (2000) . . . . . . . . . 40
2.23 Example of NN classification in a 2D feature space . . . . . . . . 41

3.1 Beating in the time domain . . . . . . . . . . . . . . . . . . . . . 48

3.2 Beating in the frequency domain . . . . . . . . . . . . . . . . . . 49
3.3 Onset detection example . . . . . . . . . . . . . . . . . . . . . . . 53

4.1 Maximum likelihood from Noll (1969) . . . . . . . . . . . . . . . 58

4.2 Two way mismatch from Maher and Beauchamp (1994) . . . . . 59
4.3 Note and musicological models from Ryynänen and Klapuri (2004) 61
4.4 Iterative cancellation method from Klapuri (2003a) . . . . . . . . 63
4.5 Probabilistic framework from Ryynänen and Klapuri (2005) . . . 64
4.6 Overview of the joint estimation method from Yeh (2008) . . . . 65
4.7 SONIC scheme from Marolt (2004a) . . . . . . . . . . . . . . . . 67
4.8 HMM smoothed estimation from Poliner and Ellis (2007a). . . . 68
4.9 NMF example from Smaragdis and Brown (2003). . . . . . . . . 70
4.10 Modified MP algorithm from Leveau et al. (2008). . . . . . . . . 71

xi
LIST OF FIGURES

4.11 Overview of the system proposed by Goto (2000). . . . . . . . . . 73

4.12 HTC spectral model of a single source from Kameoka et al. (2007) 74
4.13 Blackboard architecture from Martin (1996) . . . . . . . . . . . . 75
4.14 General onset detection scheme . . . . . . . . . . . . . . . . . . . 78

5.1 One semitone filter bank . . . . . . . . . . . . . . . . . . . . . . . 84

5.2 Example of the onset detection function for a piano melody . . . 86
5.3 Onset detection function for a polyphonic violin song . . . . . . . 88
5.4 SFS screenshot . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.5 Onset detection precision and recall . . . . . . . . . . . . . . . . 90
5.6 D2K itinerary of the onset detection system . . . . . . . . . . . . 94
5.7 MIREX onset detection results respect to θ . . . . . . . . . . . . 97

6.1 Binary digital piano-roll (BDP) . . . . . . . . . . . . . . . . . . . 104

6.2 TDNN architecture and data supplied during training . . . . . . 105
6.3 Sigmoid transfer function . . . . . . . . . . . . . . . . . . . . . . 105
6.4 Sinusoidal waveform and spectrum . . . . . . . . . . . . . . . . . 107
6.5 Sawtooth waveform and spectrum . . . . . . . . . . . . . . . . . 107
6.6 Clarinet waveform and spectrum . . . . . . . . . . . . . . . . . . 108
6.7 Hammond waveform and spectrum . . . . . . . . . . . . . . . . . 108
6.8 TDNN recognition accuracy as a function of pitch . . . . . . . . 110
6.9 TDDN accuracy respect to the amount of pitches . . . . . . . . . 111
6.10 Example TDNN detection using a clarinet sound . . . . . . . . . 112
6.11 Event detection accuracy using Ap for the sinusoidal timbre . . . 114
6.12 Event detection accuracy using A0p for the sinusoidal timbre . . . 115
6.13 Event detection accuracy using A00p for the sinusoidal timbre . . . 115

7.1 Iterative cancellation scheme. . . . . . . . . . . . . . . . . . . . . 121

7.2 SLM example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
7.3 Candidate cancellation example . . . . . . . . . . . . . . . . . . . 125
7.4 Interpolation example . . . . . . . . . . . . . . . . . . . . . . . . 130
7.5 Spectral smoothness measure example . . . . . . . . . . . . . . . 132
7.6 Example of combinations merged across adjacent frames . . . . . 135
7.7 Example of candidate intensities for an oboe sound . . . . . . . . 136
7.8 Example of candidate intensities for a mixture . . . . . . . . . . 137
7.9 Partial selection in the joint estimation method II . . . . . . . . 138
7.10 wDAG example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.11 SLM evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.12 Iterative cancellation accuracy adjusting the parameters . . . . . 144
7.13 Joint method I candidate identification adjusting the parameters 146
7.14 Joint method I accuracy adjusting the parameters . . . . . . . . 146
7.15 Joint method I runtime adjusting the parameters . . . . . . . . . 147

xii
 LIST OF FIGURES

7.16 Joint method II accuracy adjusting the parameters . . . . . . . . 149

7.17 Candidate identification results . . . . . . . . . . . . . . . . . . . 150
7.18 Global candidate identification results . . . . . . . . . . . . . . . 150
7.19 Pitch detection results for the iterative cancellation method . . . 151
7.20 Pitch detection results for the joint estimation method I . . . . . 151
7.21 Pitch detection results for the joint estimation method II . . . . 152
7.22 Global pitch detection results . . . . . . . . . . . . . . . . . . . . 152
7.23 Polyphony estimation results using with one source . . . . . . . . 153
7.24 Polyphony estimation results with two simultaneous sources . . . 153
7.25 Polyphony estimation results with four simultaneous sources . . . 154
7.26 Polyphony estimation results with six simultaneous sources . . . 154
7.27 Global polyphony estimation results . . . . . . . . . . . . . . . . 155
7.28 Results of the iterative cancellation method respect to the pitch . 155
7.29 Results of the joint estimation method I respect to the pitch . . . 156
7.30 Results of the joint estimation method II in respect to the pitch . 156
7.31 MIREX 2007-2008 Etot . Fig. from Bay et al. (2009) . . . . . . . 165
7.32 MIREX 2007-2008 accuracy. Fig. from Bay et al. (2009) . . . . . 166
7.33 MIREX 2007-2008 note tracking F-m. Fig. from Bay et al. (2009) 167

xiii
 List of Tables

2.1 Symbols representing note durations . . . . . . . . . . . . . . . . 34

5.1 Onset detection results using the ODB database . . . . . . . . . 91

5.2 Comparison with other works using the ODB database . . . . . . 92
5.3 Onset detection results using short context . . . . . . . . . . . . 93
5.4 Overall MIREX 2009 onset detection results . . . . . . . . . . . . 95
5.5 MIREX 2009 onset detection runtimes . . . . . . . . . . . . . . . 96
5.6 MIREX 2009 onset detection results using the best θ . . . . . . . 96
5.7 MIREX 2009 poly-pitched results . . . . . . . . . . . . . . . . . . 96

6.1 Frame-by-frame and note detection accuracy using TDNN . . . . 109

6.2 Frame level cross-detection results using TDNN . . . . . . . . . . 113
6.3 Note level cross-detection results using TDNN . . . . . . . . . . . 113
6.4 Event accuracy for each activation function using kNN . . . . . . 116
6.5 Note accuracy for each activation function using kNN . . . . . . 116

7.1 Iterative cancellation method parameters . . . . . . . . . . . . . 142

7.2 Joint estimation method I parameters . . . . . . . . . . . . . . . 145
7.3 Joint estimation method II parameters . . . . . . . . . . . . . . . 148
7.4 MIREX (2007) note tracking runtimes . . . . . . . . . . . . . . . 158
7.5 MIREX (2007) note tracking results based on onset and pitch . . 159
7.6 MIREX (2007) frame by frame evaluation results . . . . . . . . . 160
7.7 MIREX (2007) frame by frame runtimes . . . . . . . . . . . . . . 161
7.8 MIREX (2008) frame by frame runtimes . . . . . . . . . . . . . . 162
7.9 MIREX (2008) note tracking runtimes . . . . . . . . . . . . . . . 162
7.10 MIREX (2008) frame by frame evaluation results . . . . . . . . . 163
7.11 MIREX (2008) note tracking results . . . . . . . . . . . . . . . . 164

xv
 Introduction
1
Automatic music transcription is a music information retrieval (MIR) task which
involves many different disciplines, such as audio signal processing, machine
learning, computer science, psychoacoustics and music perception, music theory,
and music cognition.
The goal of automatic music transcription is to extract a human readable
and interpretable representation, like a musical score, from an audio signal. A
score is a guide to perform a piece of music, and it can be represented in different
ways. The most extended score representation is the modern notation used in
Western tonal music. In order to extract a readable score from a signal, it is
necessary to estimate the pitches, onset times and durations of the notes, the
tempo, the meter and the tonality of a musical piece.
The most obvious application of automatic music transcription is to help
a musician to write down the music notation of a performance from an audio
recording, which is a time consuming task when it is done by hand. Besides this
application, automatic music transcription can also be useful for other MIR
tasks, like plagiarism detection, artist identification, genre classification, and
composition assistance by changing the instrumentation, the arrangement or
the loudness before resynthesizing new pieces. In general, music transcription
methods can also provide information about the notes to symbolic music
algorithms.
The transcription process can be separated into two main stages: to convert
an audio signal to a piano-roll representation, and to convert the estimated
piano-roll into musical notation.
As pointed out by Cemgil et al. (2003), most authors only consider automatic
music transcription as an audio to piano-roll conversion, whereas piano-roll to
score notation can be seen as a separate problem. This can be justified since
the processes involved in audio to piano-roll notation include pitch estimation
and temporal note segmentation, which constitutes a challenging task itself.
The piano-roll to score process involves tasks like tempo estimation, rhythm

1
1. INTRODUCTION

quantization, key detection or pitch spelling. This stage is more related to the
generation of human readable notation.
In general, a music transcription system can not obtain the exact score
that the musician originally read. Musical audio signals are often expressive
performances, rather than simple mechanical translations of notes read on a
sheet. A particular score can be performed by a musician in different ways.
As scores are only guides for the performers to play musical pieces, converting
the notes present in an audio signal into staff notation is an ill-posed problem
without a unique solution.
However, the conversion of a musical audio signal into a piano-roll represen-
tation without rhythmic information only depends on the waveform. Rather
than a score-oriented representation, a piano-roll can be seen as a sound-
oriented representation which displays all the notes that are playing at each
time. The conversion from an audio file into a piano-roll representation is done
by a multiple fundamental frequency (f0 ) estimation method. This is the main
module of a music transcription system, as it estimates the number of notes
sounding at each time and their pitches.
For converting a piano-roll into a readable score, other harmonic and
rhythmic components must also be taken into account. The tonality is related
with the musical harmony, showing hierarchical pitch relationships based on a
key tonic. Source separation and timbre classification can be used to identify the
different instruments present in the signal, allowing the extraction of individual
scores for each instrument. The metrical structure refers to the hierarchical
temporal structure. It specifies how many beats are in each measure and what
note value constitutes one beat, so bars can be added to the score to make it
readable by a musician. The tempo is a measure to specify how fast or slow is
a musical piece.
A music transcription example is shown in Fig. 1.1. The audio performance
of the score in the top of the figure was synthesized for simplification, and
it did contain neither temporal deviations nor pedal sustains. The piano-roll
inference was done without errors. The key, tempo, and meter estimates can be
inferred from the waveform or from the symbolic piano-roll representation. In
this example, these estimates were also correct (except from the anacrusis1 at
the beginning, which causes the shift of all the bars). However, it can be seen
that the resulting score differs from the original one.
When a musician performs a score, the problem is even more challenging,
as there are frequent temporal deviations, and the onset and duration of the
notes must to be adjusted (quantized) to obtain a readable score. Note that
quantizing temporal deviations implies that the synthesized waveform of the
1 The term anacrusis refers to the note or sequence of notes which precede the beginning

of the first bar.

2
 A M.me Marie Pleyel
Nocturne
Frédéric Chopin

Op. 9 N. 2

!
Andante ( " =132) #*
12121

% !# ! !3 !4 !3 #
! ! !5 !$ ! !2 ! # !3 !2 !1 # !3 ! !2 #
54 5 4

' % % 12 !#
8 !&
1
& &
1 2 1 1

&
!
espress. dolce & &
!
! !! ! !! ! !!! ! !!! $ ! !!! ! !!! % ! $ !!! ! !!! !
$ ! !! !! !!! ! !!! % ! % !!!
( % % 12 ) ! ! !" ! %
Piano

! ! ! " ! ! ! ! !" ! $ !" !

% 8 ! ! " !" !" $ !"
" " !
" !
" !. " " " !" " ! . " ! !. " !. "
!. !. !. !. !. " " !. " !. "
!.
!5 !4 !3 ! !4 3 1 2 1 + 1%! !
! !!!# ! # !3Performer % !+ ! ! !, ! ! ! ! $ ! ! % !+ ! !+ $ ! ! ! ! % ! # ! !
! !, $ ! ! ! !3" interpretation
3 4 3
% 1
54 2 4 2 1

'%%! - !
4 1

& &
!
&
! !! % $ !! !
f %! ! ! !!! ! !!! ! ! ! ! ! !! ! !!
p !
! ! $ ! !!
cresc.
! ! "! ! ! !" ! % !! !" ! ! !" ! ! " ! ! " ! ! !! ! !! !!
( %% " ! ! ! ! !" !
! !
%! !" !"
" ! " !" " !. " !. " !. " !. " " "
!. " ! .
!. !. !.

..... 3 4 !5" !4" !3" !" !1" !"4!"# 1 3

% !/ # ! ! ! !" ! !1 % !2 # ! !# ! !2 ! !4 # !5
23

'%% !# ! ! ! !# !#
7 1 1
&
!
$ ! !!! ! !!! ! !!! ! % !!! f ! % !!! ! !! ! !!! ! !!! ! !!! ! !!! ! !! ! !!
p pp
( % % !" ! ! ! % ! " ! ! ! " ! ! " ! " ! !
% $ !" !" $ !" !Multiple ! !" ! $! !"
!"
f0 estimation

!. " ! " ! . "

! "
!" "
!. " ! " !.
. !. " .
!. . "

Tempo I.

!5 #
poco ritard.
$ !4
poco rallent.

% !" !" !" !" ! ! ! # !3 !2 # !3 !1 # !!2 ## !1 # 2" " 3 4 4 3 4 5

$ !! !!" !! !! 0 $ !! $ !! % !!! $ !!!" % !!"
4 3 2 1 2 5 5

!#
4 5

'%%
4

& $!# !#
10 1

" " " "

!
1 2 2

! !!! % ! !!! ! !!! ! !!! % ! !!! ! !!! % ! !!! $ ! !!! ! !!! ! !!! !!" $ 0 !!" " " % !!" !!"
f
( %% ! ! ! !" ! ! " ! ! ! 0! $! !
"
% %! !" $ " $ ! !" ! !"
! ! !"
!"
! . !" " !. " "
!. " !. " !. " !. " !. !. "
Creative Commons Attribution-ShareAlike 3.0

Piano-roll to score conversion

! . . .
O YZ . . . . . . . 3. . . . .
& O O ` . . .. . .. . .. . ... P . ... . ... O . P .. . .. . .. . ... . ... . .
1 2

% O O Y`Z E . . . . O . . . . . . . . E . . . . . . . E . E O.
Piano
Inst 2

O . . . . O. P. P.
. . . .
. . . .
4
O P. . . . . !
5
. . . . P. . . . O. . . .
& O O O ... . . .. .
. . P .. . ... . .. .
.
. O . .. ## .
% OO E . . . . . . . . . . . . . . . O . .
. . . .
Figure 1.1: Music transcription example from Chopin (Nocturne, Op. 9, N. 2).
O .
.
. O. P. . .
6
.
O . . O. P. O. . . P. . . 7
. . P ... . . O . .
& O O .. P . .
. . .
P .. O . .. . .. . .. . ... O . ... . 3
. . . E . . . E E
% OO
O . . . .
.
.
. . . O. . P. . P.
O.

. . . . . . ! 9
O P. . . . . . . . . . . . . . . . .
P .. .. .. . .. . .. .
8 10

& O O O .. . . .. . . . .. . .
. . . . . . .. . ..
.
% O E . . . . . . . . . . . . . . E . E . . O .. .. . . . .
OO . . . . P . . . .
. .
O . P . . . .
P. . P .. .. ... .. O O .. . P .. ...
11 12

&OO . . . P. . . . .
1. INTRODUCTION

resulting score would not exactly match the original audio times. This is the
reason why the piano-roll is considered as a sound-oriented representation.
This dissertation is mainly focused on the multiple f0 estimation issue, which
is crucial for music transcription. This is a extremely challenging task which
has been addressed in several doctoral theses, such as Moorer (1975), Maher
(1989), Marolt (2002), Hainsworth (2003), Cemgil (2004), Bello (2004), Vincent
(2004), Klapuri (2004), Zhou (2006), Yeh (2008), Ryynänen (2008), and Emiya
(2008).
Most multiple f0 estimation methods are complex and have high computa-
tional costs. As discussed in chapter 3, the estimation of multiple simultaneous
pitches is a challenging task due to the number of theoretical issues.
The main contributions of this work are a set of novel efficient methods
proposed for multiple fundamental frequency estimation (chapters 6 and 7). The
proposed algorithms have been evaluated and compared with other approaches,
yielding satisfactory results.
The detection of the beginnings of musical events on audio signals, or onset
detection, is also addressed in this work. Onset times can be used for beat
tracking, for tempo estimation, and to refine the detection in a multiple f0
estimation system. A simple and efficient novel methodology for onset detection
is described in chapter 5.
The proposed methods have also been applied to other MIR tasks, like genre
classification, mood classification, and artist identification. The main idea was
to combine audio features with symbolic features extracted from transcribed
audio files, and then use a machine learning classification scheme to yield the
genre, mood or artist. These combined approaches have been published in (Lidy
et al., 2009, 2007, 2008) and they are beyond the scope of this PhD, which is
mainly focused on music transcription itself.
This work is organized as follows. The introductory chapters 2, 3, and 4
describe respectively the theoretical background, the multiple f0 problem, and
the state of the art for automatic music transcription. Then, novel contributions
are proposed for onset detection (5), and multiple fundamental frequency
estimation (6, 7), followed by the overall conclusions and future work (8).

4
Outline
2 - Background. This chapter introduces the theoretical background, defining
the signal processing, music theory, and machine learning concepts that
will be used in the scope of this work.

3 - Music transcription. The multiple f0 estimation problem and the re-

lated theoretical issues are described in this chapter, followed by an
introduction to the onset detection problem.

4 - State of the art. This chapter presents an overview of the previous

approaches for single f0 estimation, multiple f0 estimation, and onset
detection. The review is mainly focused on the multiple f0 estimation
methods, proposing a novel categorization of the existing approaches.

5 - Onset detection using a harmonic filter bank. A novel onset de-

tection method based on the properties of harmonic musical sounds is
presented, evaluated and compared with other works.

6 - Multiple pitch estimation using supervised learning methods.

Novel supervised learning methods are proposed in a simplified scenario,
considering synthesized instruments with constant temporal envelopes.
For this task, neural networks and nearest neighbors methods have been
evaluated and compared.

7 - Multiple f0 estimation using signal processing methods. Efficient

iterative cancellation and joint estimation methods to transcribe real
music are proposed in this chapter. These methods have been evaluated
and compared with other works.

8 - Conclusions and future work. The conclusions and future work are
discussed in this chapter.

5
 Background
2
This chapter describes the signal processing, music theory, and machine learning
concepts needed to understand the basis of this work.
Different techniques for the analysis of audio signals based on the Fourier
transform are first introduced. The properties of musical sounds are presented,
classifying instruments according to their method of sound production and
to their spectral characteristics. Music theory concepts are also addressed,
describing the harmonic and temporal structures of Western music, and how can
it be represented using written and computer notations. Finally, the machine
learning techniques used in this work (neural networks and nearest neighbors)
are also described.

2.1 Analysis of audio signals

A signal is a physical quantity that is function of one or more independent
variables such as time, distance, or pressure. Sounds are air pressure signals
which frequencies are in the range that humans can hear (approximately, from
20 to 20,000 Hz1 ). The variation of the air pressure amplitude as a function of
time within this range is called audio waveform.
A waveform can be modulated into a physical medium as it happens for
a magnetic tape. The physical properties of an original sound can also be
converted into a sequence of numbers that can be stored in digital form as in a
CD.
The accuracy of the conversion of an analog audio waveform x(t) into a
digital audio waveform x[n], depends on the sampling rate fs , which determines
how often the sound is sampled. The sampling depth is the maximum numerical
size of each sampled value, which is usually expressed as the number of bits
involved in coding the samples.
1 In the International System of Units, the unit of frequency is the Hertz. 1 Hz means that

an event repeats once per second.

7
2. BACKGROUND

Systems for automatic music transcription are usually implemented using

computers, so the audio signals to be analyzed are in digital form. A digital
audio waveform has some drawbacks respect to the analog representation, like
aliasing or quantization noise, but these negative effects can be partially reduced
using high sampling rates. The most commonly used sampling rate for audio
signals is fs = 44, 100 Hz, although fs = 22, 050 Hz can be sufficient for certain
tasks.

2.1.1 Fourier transform

The information of the discrete waveform can be used directly, for example, to
detect periodicities in monophonic2 sources, by searching for a repetitive pattern
in the signal. The waveform also provides information about the temporal
envelope, that can be used for some tasks such as beat detection. However, the
time domain data is not practical for some approaches that require a different
kind of information.
A waveform can be analyzed using the Fourier transform (FT) to map it
into the frequency domain. The FT performs the decomposition of a function
in a sum of sinusoids of different frequencies, showing the signal within each
given frequency band over a range of frequencies. It is a widely used technique
for frequency analysis tasks.
The standard Fourier transform (see Eq. 2.1) is well defined for continuous
pure sine waves with infinite length.
Z +∞  
FTx (f ) = X(f ) = x(t)e−j2πf t dt 2.1 
−∞

The Fourier transform for discrete signals (discrete Fourier transform) is

defined in Eq. 2.2. The sequence of complex numbers x(n) is transformed into
a sequence of complex numbers X(k). Each spectral bin k is a frequency which
value depends on the sampling rate used (see below).

+∞
X  
DFTx (k) = X(k) = x(n)e−j2πkn 2.2 
n=−∞

In real world, signals have finite length. X[k] is defined in Eq. 2.3 for a
discrete finite signal x[n].

N
X −1
2π
 
DFTx [k] = X[k] = x[n]e−j N kn , k = 0, . . . , N − 1 2.3 
n=0

2 Only one note playing simultaneously.

8
 2.1. ANALYSIS OF AUDIO SIGNALS

Im(z )

b z = a + jb
!
|z| = a2 + b2
φ(z) = arctan(b/a)
a Re(z )

Figure 2.1: Complex plane diagram. Magnitude and phase of the complex
number z are shown.

The Shannon theorem limits the number of useful frequencies of the discrete
Fourier transform to the Nyquist frequency (fs /2). The frequency of each
spectral bin k can be easily computed as fk = k(fs /N ) since the N bins are
equally distributed in the frequency domain of the transformed space. Therefore,
the frequency resolution of the DFT is ∆f = fs /N .
The equations above are described in terms of complex exponentials. The
Fourier transform can also be expressed as a combination of sine and cosine
functions, equivalent to the complex representation by the Euler’s formula.
If the number of samples N is a power of two, then the DFT can be efficiently
computed using a fast Fourier transform (FFT) algorithm. Usually, software
packages that compute the FFT, like FFTW3 from Frigo and Johnson (2005),
use Eq. 2.3, yielding an array of complex numbers.
Using complex exponentials, the radial position or magnitude |z|, and the
angular position or phase φ(z) can easily be obtained from the complex value
z = a + jb (see Fig. 2.1).
The energy spectral density (ESD) is the squared magnitude of the DFT of
a signal x[n]. It is often called simply the spectrum of a signal. A spectrum
can be represented as a two-dimensional diagram showing the energy of a signal
|X[k]|2 as a function of frequency (see Fig. 2.2). In the scope of this work, it will
be referred as power spectrum (PS), whereas magnitude spectrum (MS) will be
referred represent the DFT magnitudes |X[k]| as a function of frequency.
Spectra are usually plotted with linear amplitude and linear frequency scales,
but they can also be represented using a logarithmic scale for amplitude,
frequency or both. A logarithmic magnitude widely used to represent the
amplitudes is the decibel.
 
dB(|X[k]|) = 20 log(|X[k]|) = 10 log(|X[k]|2 ) 2.4 
3 Fastest Fourier Transform in the West. http://www.fftw.org

9
2. BACKGROUND

0.3 5000
2
x[n] 4500 |X[k]|
0.2

4000

0.1
3500

3000
0
DFT 2500
-0.1
2000

-0.2 1500

1000
-0.3

500

-0.4
0 500 1000 1500 2000 2500 3000 3500 4000 0
0 50 100 150 200 250 300 350 400 450 500

Figure 2.2: Power spectrum of a piano waveform excerpt (I-011PFNOM.wav

from Goto (2003) RWC database).

In general, the perceived magnitude of physical variables does not directly

correspond with a physical measure. Loudness can be defined as the subjective
judgment of the intensity of a sound, and it is the correlate of the physical
amplitude. The relation between the physical amplitude and the perceived
loudness of a signal is not trivial, and it has been widely studied (Moore, 1997).
Decibels are related to the human perception of amplitude, but true loudness
is a subjective measure which varies from person to person. There exist measures
such as A-weighting, based on the work by Fletcher and Munson (1933) to
determine loudness level contours for various sound levels, that attempt to get
a loudness measure as perceived by an average listener. Sone and Phon are
units used to measure the loudness.

Fourier transform limitations

The analysis of discrete and finite signals presents some limitations. First, the
continuous to discrete conversion process can produce aliasing and quantization
noise. The solution to the aliasing problem is to ensure that the sampling rate
is high enough to avoid any spectral overlap or to use an anti-aliasing filter.
The DFT also introduces drawbacks like spectral leakage and the picket
fence effect. Spectral leakage is an effect where, due to the finite nature of the
analyzed signal, small amounts of energy are observed in frequency components
that do not exist in the original waveform, forming a series of lobes in the
frequency domain.
The picket fence is an effect related to the discrete nature of the DFT
spectrum, which is analogous to looking at it through a sort of picket fence,
since we can observe the exact behavior only at discrete points. Therefore,
there may be peaks in a DFT spectrum that will be measured too low in level,
and valleys that will be measured too high, and the true frequencies where the
peaks and valleys are will not be exactly those indicated in the spectrum.

10
 2.1. ANALYSIS OF AUDIO SIGNALS

This effect is unavoidable since the computation of the spectrum is limited

to integer multiples of the frequency resolution. However, the hidden points
can be estimated using interpolation or zero padding. These techniques can not
increment frequency resolution, but they allow to estimate the amplitude and
frequency of the hidden points.
The direct truncation of the waveform samples (also known as rectangular
windowing) leads to undesirable effects like high spectral leakage. To minimize
this effect, it is convenient to multiply the samples in the frame by a smooth
window shape to remove abrupt edges. Commonly used windows are Hanning,
Hamming, Blackman or Blackman-Harris. They have shapes that are generally
positive functions, symmetric, and bell shaped. The tradeoffs of using different
window functions are compared and discussed by Harris (1978).
Zero padding consists on appending samples with zero values at the end
of the input frame in the time domain before doing the DFT. This technique
does not increase spectral resolution, but merely interpolates the values of the
spectrum at more values.
Usually, in spectral analysis, direct interpolation does not get a better
estimation of the hidden points than zero padding. There are different
interpolation methods, like linear, quadratic or cubic, with increasing accuracy
and cost. Choosing one of these methods depends on the accuracy required, but
computational cost must also be considered.

2.1.2 Time-frequency representations

In some applications, it is convenient to represent the spectral information as a
function of time. Although there exists only one Fourier transform for a signal,
there are several possible time-frequency representations (TFRs).

Short time fourier transform

The most used TFR is the short time Fourier transform (STFT), which
represents the Fourier transform of successive signal frames. In the discrete
version, the input data is divided into time frames which usually overlap each
other. Each input frame is multiplied by a window, then Fourier transformed,
and the complex result is added to a matrix, which stores magnitude and phase
for each point in time and frequency. Discrete finite STFT can be expressed as:

N
X −1
2π
 
STFTw
x [k, m] = x[n]w[n − mI]e−j N kn , k = 0, . . . , N − 1 2.5 
n=0

where w is the window function, m is the window position index, and I is the
hop size.

11
 "3dspectrogram_piano_011.txt" matrix
2. BACKGROUND

70
60
70
50
60
40
50
30
|X[k]| 40 20
30
10
20
0
10
0
60
50
40
30
00
20 k
11 2
2 3 10
3 4
t 4 55 0

Figure 2.3: Magnitude spectrogram for the beginning section of a piano note.
Only the first 60 spectral bins and the first 5 frames are shown. Spectrum at
each frame is projected into a plane.

The hop size of the STFT determines how much the analysis starting time
advances from frame to frame. Like the frame length (window size), the choice
of the hop size depends on the purposes of the analysis. In general, a small hop
produces more analysis points and therefore, smoother results across time, but
the computational cost is proportionately increased.

Choosing a short frame duration in the STFT leads to a good time resolution
and a bad frequency resolution, and a long frame duration results in a good
frequency resolution but a bad time resolution. Time and frequency resolutions
are conjugate magnitudes, which means that ∆f ∝ 1/∆t, therefore they can
not simultaneously have an arbitrary precision. The decision about the length
of the frames in the STFT to get an appropriate balance between temporal and
frequency resolution depends on the application.

Spectrograms are three-dimensional diagrams showing the squared magni-

tude of the STFT evolving in time. Usually, spectrograms are projected into a
two-dimensional space (see the lower plane in Fig. 2.3), using colors or grey levels
to represent the magnitudes. In the scope of this work, the term magnitude
spectrogram will be referred to describe a magnitude spectrum as it changes
over time.

12
 2.1. ANALYSIS OF AUDIO SIGNALS

Figure 2.4: Wavelet filter bank with 3 levels.

Discrete wavelet transform

The discrete wavelet transform (DWT) is an alternative to the STFT. It was

introduced in order to overcome the limited time-frequency localization of the
Fourier transform for non-stationary signals.
Unlike the STFT, the DWT has a variable time-frequency resolution grid
with high frequency resolution and low time resolution in the lower frequency
area, and high temporal resolution and low frequency resolution on the higher
frequency area (see Fig. 2.5(b)). As pointed out by Tzanetakis et al. (2001),
the human ear exhibits similar time-frequency resolution characteristics.
The DWT can be generated through an algorithm called multiresolution
analysis (MRA), related to sub-band coding, by passing the input signal x[n]
through a filter bank (see Fig. 2.4). First, the samples are passed through a low
pass filter with impulse response g[n], resulting in a convolution:

∞
X  
y[n] = (x ∗ g)[n] = x[k]g[n − k] 2.6 
k=−∞

The signal is simultaneously decomposed using a high-pass filter h. The

filtered outputs are the wavelet coefficients. It is important that the two filters
are related to each other (a quadrature mirror filter must be used). Since half
the frequencies of the signal are removed, half the samples can be discarded
according to Nyquist theorem, therefore the filter outputs must be subsampled
by 2, being (y ↓ k)[n] = y[kn] the subsampling operation.
A variety of different wavelet transforms have been proposed in the literature,
like Haar (1911) or Daubechies (1988) transforms. One of the main drawbacks
of the DWT respect to the STFT is the higher computational cost. Extensive
reviews of wavelets have been done by Daubechies (1992) and Mallat (1999).
When the mother wavelet is a windowed sinusoid, the wavelet transform can
be interpreted as a constant Q Fourier transform.

13
2. BACKGROUND

∆f
∆f

∆t ∆t
(a) STFT (b) DWT

∆f ∆f

∆t ∆t
(c) Constant Q (d) Filter bank

Figure 2.5: Time-frequency resolution grids without overlap for STFT, DWT,
constant Q transform from Brown (1991), and a filter bank with 6 bands.

Constant Q transform

Using the Fourier transform, all the spectral bins obtained are equally spaced
by a constant ratio ∆f = fs /N . However, the frequencies of the musical notes
(see section 2.3) are geometrically spaced in a logarithmic scale4 .
The constant Q transform is a calculation similar to the Fourier transform,
but with a constant ratio of frequency to resolution Q. This means that each
spectral component k is separated by a variable frequency resolution ∆fk =
fk /Q.
Brown (1991) proposed a constant Q transform in which the center
frequencies fk can be specified as fk = (2k/b )fmin , where b is the number of
filters per octave and fmin is the minimum central frequency considered. The
transform using Q = 34 is similar (although not equivalent) to a 1/24 octave
filter bank. The constant Q transform for the k-th spectral component is:

1
N [k]−1
2π  
w[k, n]x[n]e−j N [k] Qn
X
XQ [k] = 2.7 
N [k] n=0

4 This scale is also related to the human frequency perception.

14
 2.1. ANALYSIS OF AUDIO SIGNALS

frequency

b1 ... bi ... bB Energy in

each band

Figure 2.6: Example of a filter bank with triangular shaped bands arranged in
a logarithmic frequency scale.

where N [k] is the window size (in samples) used to compute the transform of
the frequency k:
 
N [k] = fs /∆fk = (fs /fk )Q 2.8 
The window function w[k, n] used to minimize spectral leakage has the same
shape but a different length for each component. An efficient implementation
of the constant Q transform was described by Brown and Puckette (1992).
The main drawback with this method is that it does not take advantage of
the greater time resolution that can be obtained using shorter windows at high
frequencies, loosing coverage in the time-frequency plane (see Fig. 2.5(c)).

2.1.3 Filters in the frequency domain

In applications where some frequencies are more important than others, it is

useful to isolate certain frequency components. Filters in the frequency domain
(see Fig. 2.5(d)) serve to provide information about defined frequency regions,
and they can be used to enhance wanted frequency components or to remove
unwanted ones.
A filter bank separates the input signal into several components using an
array of band pass filters. Each component carries a single frequency subband
of the original signal. Therefore, the output of a filter bank is an array of filtered
values, each corresponding to the result of filtering the input spectrum through
an individual filter (see Fig. 2.6).
Most filter banks have filters with their edges placed so that they coincide
with the center frequencies of adjacent filters. For each filter, frequency
components within its pass band are weighted by the magnitude response of
the i-th filter |Hi [k]|, then squared and summed, as shown in Eq. 2.9.

15
2. BACKGROUND

K−1
X  
bi = (|X[k]| · |Hi [k]|)2 2.9 
k=0

Perceptually motivated scales

Psychoacoustic scales have been constructed to imitate the frequency resolution

of human hearing. A widely used psychoacoustic scale is the Mel scale
introduced by Stevens et al. (1937). A Mel frequency (see Fig. 2.7) is related to
the linear frequency through this relation:

f
   
Mel(f ) = 2595 log +1 2.10 
700
As Huang et al. (2001) points, one Mel represents one-thousandth of the
pitch of 1 kHz, and a doubling of Mels produces a perceptual doubling of pitch.
Other psychoacoustic scale is the Bark introduced by Zwicker et al. (1957),
which partitions the hearing bandwidth into perceptually equal frequency bands
(critical bands). If the distance between two spectral components is less than
the critical bandwidth, then one masks the other.
The Bark scale, also called critical band rate (CBR), is defined so that the
critical bands of human hearing have a width of one bark. This partitioning,
based on the results of psychoacoustic experiments, simulates the spectral
analysis performed by the basilar membrane, in such a way that each point on
the basilar membrane can be considered as a bandpass filter having a bandwidth
equal to one critical bandwidth, or one bark.
A CBR filter bank is composed of a set of critical band filters, each one
corresponding to one bark. The center frequencies to build the filter bank
are described by Zwicker (1961). The Mel filter bank is composed of a set of
filters with a triangular shape and equally spaced in terms of Mel frequencies.
Shannon and Paliwal (2003) showed that Bark and Mel filter banks have similar
performance in speech recognition tasks.
The Mel frequency cepstral coefficients (MFCC) have been extensively used
in tasks such as automatic speech recognition and music processing. To compute
the MFCC features, the power spectrum of the signal is first computed and
apportioned through a Mel filter bank. The logarithm of the energy for each
filter is calculated before applying a Discrete Cosine Transform (DCT, see
Ahmed et al., 1974) to produce the MFCC feature vector. The DCT of a
discrete signal x[n] with a length N is defined as:

N
X −1
π
 
1
  
DCTx [i] = x[n] cos i n+ 2.11 
n=0
N 2

16
 2.2. ANALYSIS OF MUSICAL SIGNALS

3500

3000

2500

2000

Mel scale

1500

1000

500

0
0 2000 4000 6000 8000 10000
Hertz scale

Figure 2.7: Mel scale and linear frequency.

The MFCC are the obtained DCT amplitudes. In most applications, the
dimensionality of the MFCC representation is usually reduced by selecting only
certain coefficients.
The bandwidth of a filter can be expressed using an equivalent rectangular
bandwidth (ERB) measure. The ERB of a filter is defined as the bandwidth of
a perfectly rectangular filter with a unity magnitude response and same area as
that filter. According to Moore (1995), the ERB bandwidths bc of the auditory
filter at the channel c obey this equation:
 
bc = 0.108fc + 24.7 Hz 2.12
 
being fc the center frequency of the filter.
A filter with a triangular shape can be useful for some applications, but
other shapes are needed to model the auditory responses. Filter frequency
responses can be expressed in terms of a gaussian function (Patterson, 1982), a
rounded exponential (Patterson et al., 1982), and a gammatone or “Patterson-
Holdsworth” filter (Patterson et al., 1995). Gammatone filters are frequently
used in music analysis, and a description of their design and implementation
can be found in (Slaney, 1993). Auditory filter banks have been used to model
the cochlear processing, using a set of gammatone filters uniformly distributed
in the critical-band scale.

2.2 Analysis of musical signals

Musical signals are a subset of audio signals, and they have particular
characteristics that can be taken into account for their analysis. In this section,
some temporal and spectral features of musical sounds are described.

17
2. BACKGROUND

0.04
RMS

0.035

0.03

0.025

0.02

0.015

0.01

0.005

0
0 20 40 60 80 100 120

Attack Sustain Release

Figure 2.8: Waveform of a flute sound (excerpt of I-331FLNOM from Goto

(2003), RWC database) with fs = 44, 100 Hz, and RMS levels using N = 1024.

2.2.1 Dynamics
Musical instruments produce sounds that evolve in time. The beginning of a
sound is known as the onset time, and its temporal end is the offset time. The
amplitude envelope refers to a temporally smoothed curve of the sound intensity
as a function of time, which evolves from the onset to the offset times.
The envelope of a signal is usually calculated in the time domain by lowpass
filtering (with a 30 Hz cut-off frequency) the root mean square (RMS) levels of
a signal. The RMS levels E[n] can be obtained as:
v
 
u N −1
u1 X
E[n] = t x2 [n + i] 2.13 
N i=0

where N is the size of the frame. Real sounds have a temporal envelope with an
attack and release stages (like percussion or plucked strings), or attack, sustain
and decay segments (like woodwind instruments)5 . The automatic estimation of
the intra-note segment boundaries is an open problem, and it has been addressed
by some authors like Jensen (1999), Peeters (2004), and Maestre and Gómez
(2005).

5 Synthesizers generate amplitude envelopes using attack, decay, sustain and release

(ADSR), but this segmentation is not achievable in real signals, since the decay part is often
not clearly present, and some instruments do not have defined sustain or release parts.

18
 2.2. ANALYSIS OF MUSICAL SIGNALS

The attack of a sound is formally defined as the initial interval during which
the amplitude envelope increases. For real sounds, Peeters (2004) considers
attack as the initial interval between the 20% and the 80% of the maximum
value in the signal, to take into account the possible presence of noise.
Transients are fast varying features characterized by sudden bursts of noise,
or fast changes of the local spectral content. During a transient, the signal
evolves in a relatively unpredictable way. A transient period is usually present
during the initial stage of the sound, and it often corresponds to the period
during which the instrument excitation is applied, though in some sounds a
transient can also be present in the release stage.
A vibrato is a periodic oscillation of the fundamental frequency, whereas
tremolo refers to a periodic oscillation in the signal amplitude. In both cases,
this oscillation is of subsonic frequency.

2.2.2 Timbre
In music, timbre is the quality that distinguishes musical instruments. The
American Standards Association (1960) defines timbre as that attribute of
sensation in terms of which a listener can judge that two sounds having the
same loudness and pitch are dissimilar.
From an audio analysis point of view, it is convenient to understand the
characteristics that make an instrument different from others. Timbral features
extracted from the waveform or the spectrum of an instrument are the basis for
automatic instrument classification.
It was shown by Grey (1978) and Wessel (1979) that important timbre
characteristics of the orchestral sounds are attack quality (temporal envelope),
spectral flux (evolution of the spectral distribution over time), and brightness
(spectral centroid).
The spectral flux (SF) is a measure of local spectrum change, defined as:

K−1
X  
SF(t) = (X̃t [k] − X̃t−1 [k])2 2.14 
k=0

where X̃t [k] and X̃t−1 [k] are the energy normalized Fourier spectra in the current
and previous frames, respectively:

|X[k]|  
X̃[k] = PK−1 2.15
 
k=0 |X[k]|

The spectral centroid (SC) indicates the position of the sound spectral center
of mass, and it is related to the perceptual brightness of the sound. It is
calculated as the weighted mean of the frequencies present in the signal, and
the weights are their magnitudes.

19
2. BACKGROUND

K−1
X  
SCX = k X̃[k] 2.16
 
k=0

Besides these features, there exist a number of characteristics that describe

a particular timbre. Some of them used for automatic instrument classification
are, for instance, spectral skewness, spectral kurtosis, spectral spread, spectral
irregularity, spectral roll-off, MFCC, inharmonicity, odd-to-even ratio, tristim-
ulus or temporal centroid. More information about these descriptors can be
found in the works from Peeters (2004) and Herrera et al. (2006).

2.2.3 Taxonomy of musical instruments

In the literature, different taxonomies for instrument classification have been
proposed. The most referenced is the one proposed by Hornbostel and Sachs
(1914), and updated by Sachs (1940), who included the electrophones category.
Five families of instruments are considered in this taxonomy:

1. Idiophones, which sound is primarily produced by way of the instrument

itself vibrating without the use of membranes or strings. This group
includes all percussion instruments apart from drums, and some other
instruments. Four subfamilies are considered; struck, plucked, friction
and blown idiophones. Cymbals, xylophones, and nail violins belong to
this family.

2. Membranophones, which sound is primarily produced by the vibration of

a tightly stretched membrane. This group includes all drums and kazoos.
Four subfamilies are considered; struck drums, plucked drums, friction
drums, and singing membranes. Timpani and snare drums belong to this
class.

3. Chordophones, which sound is primarily produced by the vibration of

one or more strings. This group includes string instruments and keyboard
instruments. Subfamilies are simple and composite chordophones. Pianos,
harpsichords, violins, guitars, and harps belong to this family.

4. Aerophones, which sound is primarily produced by vibrating air. This

group includes brass and woodwind instruments. Subfamilies are free
aerophones and wind and brass instruments. Oboes, saxophones, trum-
pets, and flutes belong to this category.

5. Electrophones, which sound is produced by electric means. Subfamilies

are instruments that have electric action, instruments with electric am-
plification and radioelectric instruments. Electronic organs, synthesizers
and theremins belong to this category.

20
 2.2. ANALYSIS OF MUSICAL SIGNALS

0.1 35

0.08 x[n] |X[k]|

30
0.06

25
0.04

0.02
T0 DFT
20

15
-0.02

-0.04 10

-0.06
5
-0.08

-0.1 0
0 500 1000 1500 2000 0 100 200 300 400 500

1 f0
f0 =
T0

Figure 2.9: Example waveform and spectrum of a violin excerpt (file I-

151VNNOM from Goto (2003), RWC database).

Instruments are classified in the families above depending on its exciter, the
vibrating element that transforms the energy supplied by the player into sound.
However, a complementary taxonomy can be assumed, dividing musical sounds
in two main categories: pitched and unpitched sounds.

2.2.4 Pitched musical sounds

The fundamental frequency f0 of a signal measures the number of occurrences
of a repeating event per time unit. Therefore, the fundamental frequency is only
present in sounds that are nearly periodic in the time domain. The fundamental
period T0 of a signal is the duration of one cycle in a repeating event, so it is
the reciprocal of the fundamental frequency (T0 = 1/f0 ).
Pitch is a perceptual attribute related to the fundamental frequency which
allows the order of sounds on a frequency related scale extending from low
to high (Klapuri, 2006a). More exactly, Hartmann (1996) defines pitch as the
frequency of a sine wave that is matched to the target sound by human listeners.
Pitched musical sounds are those that cause a clear pitch sensation, like a
piano or a guitar. Most chordophones and aerophones produce pitched sounds,
as some electrophones and some idiophones. Pitched musical sounds can be
divided into harmonic and inharmonic sounds.

Harmonic sounds

Many Western musical instruments, more precisely those in the chordophone

and aerophone families, produce harmonic sounds, as they are based on a
harmonic oscillator such as a string or a column of air. The spectrum of these
sounds (see Figs. 2.9 and 2.10) shows a series of overtone partials regularly
spaced (harmonics).

21
2. BACKGROUND

Figure 2.10: Example waveform and spectrogram of a piano note (file I-

011PFNOM from Goto (2003), RWC database).

In an ideal harmonic sound, the harmonics are integers multiple of a

fundamental frequency. Therefore, the frequency differences of the harmonics
corresponds approximately to the fundamental frequency, and the f0 of a
harmonic sound can be defined as the greatest common divisor of the harmonic
frequencies.
However, in real harmonic sounds, the partial overtones are usually not
exactly multiples of the fundamental partial. This effect is known as inhar-
monicity, and it appears when the frequency of a partial h is not exactly hf0 .
Although inharmonicity can be a problem for analyzing harmonic sounds, it
is not perceptually unpleasant. A slightly inharmonic spectrum adds certain
warmth into the sound.
As pointed out by Schouten (1940), the pitch of a complex tone can be
perceived even though the frequency component corresponding to the f0 may not
be present (missing fundamental). Note that, if the fundamental were necessary,
normal male speech would have no perceivable pitch over the telephone, as
frequencies below 300 Hz are generally filtered out6 .
In most string instruments, the higher order partials gradually shift upwards
in frequency. The inharmonicity of a string depends on its physical characteris-
tics, such as tension, stiffness, and length. For instance, a stiff string under low
tension (such as a bass string in a piano) exhibits a high degree of inharmonicity,
6 As pointed out by Huang et al. (2001), the fundamental frequency of speech can vary

from 40 Hz for low-pitched male voices to 600 Hz for children or high-pitched female voices.

22
 2.2. ANALYSIS OF MUSICAL SIGNALS

whereas a thinner string under higher tension (such as a treble string in a piano)
or a more flexible string (such as a nylon string used on a guitar or harp) exhibits
less inharmonicity.
According to Fletcher and Rossing (1988), the harmonic frequencies in a
piano string approximately obey this formula:
p  
fh = hf0 1 + Bh2 2.17 
A typical value of the inharmonicity factor for the middle pitch range of a
piano is B = 0.0004, which is sufficient to shift the 17th partial to the ideal
frequency of the 18th partial.
In some cases, there are short unpitched excerpts in pitched sounds, mainly
in the initial part of the signal. For instance, during the attack stage of wind
instruments, the initial breath noise is present before the pitch is perceived.
Inharmonic sounds are also produced by the clicking of the keys of a clarinet,
the scratching of the bow of violin, or the sound of the hammer of a piano
hitting the string, for instance.
The additive synthesis, that was first extensively described by Moorer (1977),
is the base of the original harmonic spectrum model, which approximates a
harmonic signal by a sum of sinusoids. A harmonic sound can be expressed as
a sum of H sinusoids with an error model :

H
X  
x[n] = Ah [n] cos(2πfh n + φh (0)) + [n] 2.18 
h=1

where Ah is the amplitude of the h-th sinusoid varying in function of time, fh

is the frequency of the sinusoid, and φh (0) is the initial phase.
Most models for parametric analysis of sounds are based on the additive
synthesis model developed by Mcaulay and Quatieri (1986) for speech signals,
who proposed a robust method for extracting the amplitudes, frequencies, and
phases of the component sine waves. The model assumes that the sinusoids are
stable partials of the sound with a slowly changing amplitude and frequency.
More precisely, instrument harmonics tend to have a linearly increasing angular
position and a constant radius.
The model proposed by Mcaulay and Quatieri (1986) was refined by Serra
(1997). Spectral modeling synthesis (SMS) is based on modeling sounds as
stable sinusoids (harmonics, also called deterministic part), plus noise (residual
component, also called stochastic part). Using this model, new sounds can be
generated (synthesized) from the analysis of a signal. The analysis procedure
detects harmonics by studying the time varying spectral characteristics of the
sound and represents them with time varying sinusoids. These partials are then

23
2. BACKGROUND

subtracted from the original sound and the remaining residual is represented as
a time varying filtered white noise component.
Recent parametric models, like the ones proposed by Verma and Meng (2000)
and Masri and Bateman (1996), extend the SMS model to consider transients.
When sharp onsets occur, the frames prior to an attack transient are similar,
and also the frames following its onset, but the central frame spanning both
regions is an average of both spectra that can be difficult to be analyzed.
Without considering noise or transients, in a very basic form, a harmonic
sound can be described with the relative amplitudes of its harmonics and their
evolution over time. This is also known as the harmonic pattern (or spectral
pattern). Considering only the spectral magnitude of the harmonics at a given
time frame, a spectral pattern p can be defined as a vector containing the
magnitude ph of each harmonic h:
 
p = {p1 , p2 , ..., ph , ..., pH } 2.19 
This partial to partial amplitude profile is also referred to as the spectrum
envelope. Adding the temporal dimension to obtain the spectral evolution in
time, a harmonic pattern can be written in matrix notation:
 
P = {p1 , p2 , ..., pt , ..., pT } 2.20 
In most musical sounds, the first harmonics contain most of the energy of
the signal, and sometimes their spectral envelope can be approximated using a
smooth curve.

Inharmonic sounds

The pitched inharmonic sounds have a period in the time domain and a pitch,
but their overtone partials are not approximately integer multiples of the f0 .
Usually, a vibrating bar is the sound source of these instruments, belonging to
the idiophones family. The most common are the marimba, vibraphone (see
Fig. 2.11), xylophone and glockenspiel.
As the analysis of inharmonic pitched sounds is complex and these instru-
ments are less commonly used, most f0 estimation systems that analyze the
signal in the frequency domain do not handle them appropriately.

2.2.5 Unpitched musical sounds

Unpitched musical sounds are those that do not produce a clear pitch sensation.
They belong to two main families of the Hornbostel and Sachs (1914) taxonomy:
membranophones and idiophones. All the membranophones family, many
idiophones and some electrophones produce unpitched sounds.

24
 2.2. ANALYSIS OF MUSICAL SIGNALS

Figure 2.11: Example waveform and spectrogram of a vibraphone (file I-

041VIHNM from Goto (2003), RWC database).

Figure 2.12: Example waveform and spectrogram of an unpitched sound (snare

drum, file I-421SD3N3 from Goto (2003), RWC database).

25
2. BACKGROUND

Most of these sounds are characterized by a sharp attack stage, that usually
shows a broad frequency dispersion (see Fig. 2.12). Interestingly, Fitzgerald and
Paulus (2006) comment that although synthetic7 drum sounds tend to mimic
real drums, their spectral characteristics differ considerably from those in real
drums.
Spectral centroid, bandwidth of the spectrum and spectral kurtosis are
features commonly used in unpitched sound classification.
The transcription of unpitched instruments is referred to the identification
of the timbre class and its onset and offset times, as no pitch is present. This
task will not be addressed in the scope of this thesis, which is mainly focused on
the transcription of pitched sounds. For a review of this topic, see (FitzGerald,
2004) and (Fitzgerald and Paulus, 2006).

2.2.6 Singing sounds

According to Deutsch (1998) p.172, singing sounds are produced by the human
vocal organ, which consists of three basic components: the respiratory system,
the vocal folds and the vocal tract. The respiratory system provides an excess
pressure of air in the lungs. The vocal folds chop the airstream from the
lungs into a sequence of quasi-periodic air pulses, producing a sound with a
fundamental frequency. Finally, the vocal tract modifies the spectral shape and
determines the timbre of the voice. The Fig. 2.13 shows a voice spectrogram
example.
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 (Ryynänen, 2006). In a simplified scenario, the amplitudes of
the overtone partials can be expected to decrease by about 12 dB per octave
(Sundberg, 1987). The phonation frequencies range from around 100 Hz for
male singers over 1 kHz for female singers.
The vocal tract functions as a resonating filter which emphasizes certain
frequencies called the formant frequencies. The two lowest formants contribute
to the identification of the vowel, and the higher formants to the personal voice
timbre.

2.3 Music background

The characteristics of isolated musical audio sounds have been described from a
signal processing point of view. However, music is produced by a combination
of pitched and/or unpitched overlapped signals beginning at different times. In
music, this combination of notes is not random. In most situations, a music
7 Sounds that are not generated by any real musical instrument.

26
 2.3. MUSIC BACKGROUND

Figure 2.13: Example waveform and spectrogram of a singing male voice, vowel
A (file I-471TNA1M from Goto (2003), RWC database).

piece follows basic melodic, harmonic and rhythmic rules to be pleasing to most
listeners.
In this section, some terms related to the music structure in time and
frequency are described, followed by a brief explanation for understanding a
musical score and its symbols. Different score formats and representations
commonly used in computer music are also introduced.

2.3.1 Tonal structure

Harmony is a term which denotes the formation and relationships of simulta-

neous notes, called chords, and over time, chordal progressions. A melody is
a sequence of pitched sounds with musically meaningful pitches and a metrical
structure. Therefore, the term melody refers to a sequence of pitches, whereas
harmony refers to the combination of simultaneous pitches.
A musical interval can be defined as a ratio between two pitches. The term
harmonic interval refers to the pitch relationship between simultaneous notes,
whereas melodic interval refers to the pitch interval of two consecutive notes. In
Western tonal music, intervals that cause two notes to share harmonic positions
in the spectrum, or consonant intervals, are more frequent that those without
harmonic relationships, or dissonant intervals.

27
2. BACKGROUND

C! D! F! G! A!
D! E! G! A! B!

C D E F G A B

Figure 2.14: Western note names in a piano keyboard. Only one octave is
labeled.

Musical temperaments

In terms of frequency, musical intervals are relations described by the ratio

between the respective frequencies of the involved notes. The octave is the
simplest interval in music, after the unison8 . Two notes separated by one octave
have a frequency ratio of 2:1. The human ear tends to hear two notes an octave
apart as being essentially the same. This is the reason why, in most musical
cultures (like Western, Arabic, Chinese, and Indian music), the wide range of
pitches is arranged across octaves in a logarithmic frequency scale.
Music is based on the octave relationship, but there exist different ways for
arranging a number of musical notes within an octave and assigning them a given
frequency. In Western music, the most common tuning system is the twelve tone
equal temperament, which divides each octave into 12 logarithmically equal
parts, or semitones.
In this musical temperament, each semitone is equal to one twelfth of an
octave. Therefore, every pair of adjacent notes has an identical frequency ratio
of 1:21/12 , or 100 cents. One tone is defined as a two semitones interval. Equal
temperament is usually tuned relative to a standard frequency for pitch A of
440 Hz9 .

Western musical pitches

A musical note can be identified using a letter (see Fig. 2.14), and an octave
number. For instance, C3 refers to the note C from the third octave. Notes
separated by an octave are given the same note name. The twelve notes in each
octave are called pitch classes. For example, the note C3 belongs to the same
pitch class than C4 .
8 An unison is an interval with a frequency ratio 1:1.
9 This frequency reference for tuning instruments was first adopted as the USA Standard
Pitch in 1925, and it was set as the modern concert pitch in May 1939. Before, a variety of
standard frequencies were used. For example, in the time of Mozart, the pitch A had a value
close to 422 Hz.

28
 2.3. MUSIC BACKGROUND

Figure 2.15: Musical major keys (uppercase), and minor keys (lowercase).
The number of alterations and the staff representation are shown. Fig. from
http://en.wikipedia.org/wiki/File:Circle_of_fifths_deluxe_4.svg.

There are 12 pitch classes, but only 7 note names (C,D,E,F,G,A,B). Each
note name is separated by one tone except F from E, and C from B, which have
a one semitone interval. This is because modern music theory is based on the
diatonic scale.

Musical scales and tonality

The diatonic scale is a seven note musical scale comprising five whole steps and
two half steps, with the pattern repeating at the octave. The major scale is a
diatonic scale which pattern of intervals in semitones is 2-2-1-2-2-2-1, starting
from a root note10 . For instance, the major diatonic scale with root note C is
built using the white keys of the piano. The natural minor scale has a pattern
of intervals 2-1-2-2-1-2-2.
In tonal music, a scale is an ordered set of notes typically used in a tonality
(also referred as key). The tonality is the harmonic center of gravity of a musical
excerpt. Intervals in the major and minor scales are consonant intervals relative
to the tonic, therefore they are more frequent within a given key context.

10 The root note is also called tonic or harmonic center

29
2. BACKGROUND

Figure 2.16: Harmonics and intervals. The first nine harmonics of middle
C. Their frequencies and nearest pitches are indicated, as well as the Western
tonal-harmonic music intervals. Fig. from Krumhansl (2004).

A musical excerpt can be arranged in a major or a minor key (see Fig. 2.15).
Major and minor keys which share the same signature are called relative.
Therefore, C major is the relative major of A minor, whereas C minor is
the relative minor of E major. The key is established by particular chord
progressions.

Consonant and dissonant intervals

As introduced above, harmonic and melodic intervals can be divided into
consonant and dissonant. Consonant intervals are those that cause harmonic
overlapping in some degree. For harmonic interference it is not required exact
frequency overlap, only approximation11 .
The perceptual dimension of consonance and dissonance is related to ratios
of frequencies. The ordering along the dimension of consonance corresponds
quite closely to the size of the integers in the ratios (Vos and Vianen, 1984).
The unison (1:1) and octave (2:1) are the most consonant intervals, followed
by the perfect intervals. Perfect intervals12 are the perfect fifth (3:2) and the
perfect fourth (4:3). The major third (5:4), minor third (6:5), major sixth (5:3),
11 Other temperaments, like the meantone temperament, make the intervals closer to their

ideal just ratios.

12 In the equal temperament, besides the unison and the octave, the interval ratios described

are approximate.

30
 2.3. MUSIC BACKGROUND

and minor sixth (8:5) are next most consonant. The least consonant intervals
in western harmony are the minor second (16:15), the major seventh (15:8) and
the tritone (45:32).
In music, consonant intervals are more frequent than dissonant intervals.
According to Kosuke et al. (2003), trained musicians find more difficult to
identify pitches of dissonant intervals than those of consonant intervals.
It is hard to separate melody from harmony in practice (Krumhansl, 2004),
but harmonic and melodic intervals are not equivalent. For example, two notes
separated by one octave play the same harmonic rule, although they are not
interchangeable in a melodic line.
The most elemental chord in harmony is the triad, which is a three note
chord with a root, a third degree (major or minor third above the root), and a
fifth degree (major or minor third above the third).

2.3.2 Rhythm
A coherent temporal structure is pleasing to most listeners. Music has a
rhythmic dimension, related to the placement of sounds at given instants and
their accents13 .
In the literature, there exist some discrepancies about the terminology used
to describe rhythm14 . An excellent study of the semantics of the terms used in
computational rhythm can be found in (Gouyon, 2008).
According to Fraisse (1998), a precise, generally accepted definition of
rhythm does not exist. However, as Honing (2001) points out, there seems
to be agreement that the metrical structure, the tempo (tactus) and the timing
are three main rhythmic concepts.
The metrical structure refers to the hierarchical temporal structure, the
tempo indicates how fast or slow is a musical piece, and the timing deviations
that occur in expressive performances are related to the temporal discrepancies
around the metrical grid.

Metrical structure

Meter is a concept related to an underlying division of time. From a perceptual

point of view, Klapuri (2003b) considers musical meter as a hierarchical
structure consisting of pulse sensations at different levels (time scales). Usually,
three main levels are considered in a metrical structure; the beat (tactus), the
measure (bar) and the tatum.
The beat, or tactus level, is the basic time unit in music. Handel (1989)
defines beat as a sense of equally spaced temporal units. It typically corresponds
13 In music, an accent is an emphasis placed on a particular note.
14 For example, in some works, the terms pulse and beat are equivalent.

31
2. BACKGROUND

Figure 2.17: Diagram of relationships between metrical levels and timing. Fig.
from Hainsworth (2003).

to the preferred human foot tapping rate (Klapuri et al., 2006), or to the dance
movements when listening to a musical piece.
A measure constitutes a temporal pattern and it is composed by a number
of beats. In Western music, rhythms are usually arranged with respect to a
time signature. The time signature (also known as meter signature) specifies
how many beats are in each measure and what note value constitutes one beat.
One beat usually corresponds to the duration of a quarter note15 (or crochet) or
an eighth note (or quaver) in musical notation. A measure is usually 2, 3, or 4
beats long (duple, triple, or quadruple), and each beat is normally divided into
2 or 3 basic subdivisions (simple, or compound). Bar division is closely related
to harmonic progressions.
Unfortunately, the perceived beat does not always correspond with the one
written in a time signature. According to Hainsworth (2003), in fast jazz music,
the beat is often felt as half note (or minim), i.e., double of his written rate,
whereas hymns are often notated with the beat given in minims, the double of
the perceived rate.
The tatum16 , first defined by Bilmes (1993), is the lowest level of the metric
musical hierarchy. It is a high frequency pulse that we keep in mind when
perceiving or performing music. An intuitive definition of the tatum proposed
by Klapuri (2003b) refers it as the shortest durational value in music that are
still more than incidentally encountered, i.e., the shortest commonly occurring
time interval. It frequently corresponds to a binary, ternary, or quaternary
subdivision of the musical beat. The duration values of the other notes, with
few exceptions, are integer multiples of the tatum. The tatum is not written
in a modern musical score, but it is a perceptual component of the metrical
structure.
15 Note durations are shown in Fig. 2.1.
16 In honor of Art Tatum.

32
 2.3. MUSIC BACKGROUND

The generative theory of tonal music (GTTM), introduced by Lerdahl and

Jackendoff (1983), is a detailed theory of musical hierarchies from a psychological
perspective, attempting to produce formal descriptions in a scientific sense. In
the GTTM, besides metrical structure, grouping structure is also considered.
The grouping structure is related to the musical surface, where listeners can
find motives, phrases and sections.

Tempo
The tempo (also referred as tactus) indicates the speed of the underlying beat.
It is usually measured in bpm (number of beats per minute), and it is inversely
proportional to the beat period. Having a beat period Tb expressed in seconds,
the tempo can be computed as T = 60/Tb . Like other rhythmic components, it
can vary along a piece of music.

Timing
Usually, when a score is performed by a musician, the onset times of the
played notes do not exactly correspond with those indicated in the score. This
temporal deviation, known as timing deviation (see Fig. 2.17) is frequent in
musical signals. It can be produced either by slight involuntary deviations in
the performance, or by deliberate expressive rhythm alterations, like swing.
In music psychology, emotion component of music is strongly associated
with music expressivity (Juslin et al., 2006). A musical piece can be performed
to produce different emotions in the listener (it can be passionate, sweet,
aggressive, humorous, etc). In the literature, there exist a variety of approaches
to map a song into a psychologically based emotion space, classifying it
according to its mood. Timing and metrical accents are frequently affected
by mood alterations in the performance.

2.3.3 Modern music notation

A score is a guide to perform a piece of music. In the history of music, a
number of different representations of music through written symbols have been
used. The most extended musical notation is the modern notation, originated
in European classical music to represent Western tonal music.
Modern notation is only intended for the representation of pitched sounds,
and it represents the music using a two dimensional space. A five line staff is
used as basis for this notation, where pitches are represented in the vertical axis,
and time in the horizontal axis.
Pitch is shown by placement of notes on the staff, over a line or between
two lines. The pitch can be modified by accidentals, like sharps (]), flats ([),
double sharps (×), double flats ([[), or naturals (\). Sharps and flats increase

33
2. BACKGROUND

Note Rest American name British name

Whole Semibreve

Half Minim

Quarter Crotchet

Eighth Quaver

Sixteenth Semiquaver

Thirty-second Demisemiquaver

Sixty-fourth Hemidemisemiquaver

Table 2.1: Symbols used to represent the most frequent note and rest durations.

or decrease the pitch by one semitone, respectively. Notes with a pitch outside
of the range of the five line staff can be represented using ledger lines, which
provide a single note with additional lines and spaces.
Duration is shown with different note figures (see Fig. 2.1), and additional
symbols such as dots (·) and ties (^). Notation is read from left to right.
A staff begins with a clef, which indicates the pitch of the written notes.
Following the clef, the key signature indicates the key by specifying certain
notes to be flat or sharp throughout the piece, unless otherwise indicated.
The time signature appears after the key signature. Measures (bars) divide
the piece into regular groupings of beats, and the time signatures specify those
groupings.
Directions to the performer regarding tempo and dynamics are added above
or below the staff. In written notation, the term dynamics usually refers to the
intensity of the notes17 . The two basic dynamic indications in music are p or
piano, meaning soft, and f or forte, meaning loud or strong.
In modern notation, lyrics can be written for vocal music. Besides this
notation, others can be used to represent unpitched instruments (percussion
notation) or chord progressions (e.g., tablatures).

2.3.4 Computer music notation

A digital score contains symbolic data which allow the easy calculation of
musical information and its manipulation. In a computer, a score can be stored
17 Although the term dynamics is sometimes used to refer other aspects of the execution of

a given piece, like staccato, legato, etc.

34
 2.3. MUSIC BACKGROUND

Figure 2.18: Excerpt of a musical score in modern notation. Symbols that do

not represent notes are annotated.

and represented in different ways. Musical software can decode symbolic data
and represent them in modern notation. Software like sequencers can also play
musical pieces in symbolic formats using a synthesizer.

Symbolic formats

The MIDI18 (Musical Instrument Digital Interface) protocol, introduced by the

MIDI Manufacturers Association and the Japan MIDI Standards Committee in
1983, enables electronic musical instruments, computers, and other equipment
to communicate, control, and synchronize with each other. MIDI does not
generate any sound, but it can be used to control a MIDI instrument that will
produce the specified sound. Event messages such as the pitch and intensity
(velocity) of musical notes can be transmitted using this protocol. It can also
control parameters such as volume, vibrato, panning, musical key, and clock
signals to set the tempo.
In MIDI, the pitch of a note is encoded using a number (see Fig. 2.19). A
frequency f can be converted into a MIDI pitch number n using Eq. 2.21:
 
f
  
n = round 69 + 12 log2 2.21 
440
Inversely, the frequency f of a given MIDI pitch number n can be obtained
using Eq. 2.22.
n−69  
f = 440 · 2 12 2.22
 
Other important MIDI component is the intensity of a note, which is encoded
using a number in the range [0 : 127].
18 http://www.midi.org

35
2. BACKGROUND

Figure 2.19: Equal temperament system, showing their position in the staff,
frequency, note name and MIDI note number. Fig. from Joe Wolfe, University
of South Wales (UNSW), http://www.phys.unsw.edu.au/jw/notes.html.

MIDI messages, along with timing information, can be collected and stored
in a standard MIDI file (SMF). This is the most extended symbolic file
format in computer music. The SMF specification was developed by the MIDI
Manufacturers Association (MMA). Large collections of files in this format can
be found on the web.
The main limitation of MIDI is that there exist musical symbols in modern
notation that can not be explicitly encoded using this format. For example, pitch
names have a different meaning in music, but there is no difference between C]
and D[ in MIDI, as they share the same pitch number. In the literature, there
are a number of pitch spelling algorithms to assign contextually consistent letter
names to pitch numbers according to the local key context. A comparison of
these methods can be found in the Meredith and Wiggins (2005) review.
This is because MIDI is a sound-oriented code. It was designed as a protocol
to control electronic instruments that produce sound, and not initially intended
to represent musical scores. Another example of a sound related code is CSound
score format, which was also designed for the control and generation of sounds.

36
 2.3. MUSIC BACKGROUND

CSound is a software synthesizer developed by the Medialab at MIT. For

specification details of the language, see (Vercoe, 1991), and (Boulanger, 1999).
There exist other schemes, referred as musical notation codes, that overcome
the limitations of the MIDI specification for score representation. Some
examples are MusicXML19 , SCORE20 , and Lilypond21 .
A third category of musical codes are those designed for musical data
management and analysis, like Humdrum22 , and the Essen Associative Code23 .
The work from Selfbridge-Field (1997) presents an extensive review for a
variety of musical coding schemes, including MIDI.

Visual representations of SMFs

The symbolic musical information stored in a SMF can be represented in
different ways by a computer software, as it is not a symbolic code well defined
to obtain a modern score. The musical data of a SMF can be visualized as
a piano-roll or as approximate modern notation. Different sequencers provide
very similar piano-roll representations of a MIDI file, but they generate different
scores from the same SMF, as MIDI data must be interpreted to be translated
into a score.
The piano-roll is a natural representation of a SMF with time in the
horizontal axis, and pitch in the vertical axis (see Fig. 2.20, top). Notes are
represented using horizontal bars in the time-frequency grid. Some sequencers
assign a color to the bars related to the note intensity.
Translating MIDI data into modern notation leads to quantization. Quanti-
zation is the process of aligning the musical notes to a given grid. Note onsets24
are set on their nearest beats or exact fractions of beats, depending on the
quantization unit. This process is usually done for representing a readable score
removing slight timing deviations.
Using quantization, some musical figures like triplets25 or notes shorter than
the quantization unit (like those in bar 10, Fig. 2.20) can not be represented
correctly.
More information about the problem of extracting the musical score from
a MIDI performance can be found in (Cambouropoulos, 2000). Although the
MIDI format has some limitations, as pointed out by Klapuri and Davy (2006),
standard MIDI files are usually considered a valid format for representing
symbolic music in computational music transcription.
19 http://www.recordare.com/xml.html
20 http://www.scoremus.com/score.html
21 http://lilypond.org
22 http://www.humdrum.org/Humdrum/
23 http://www.esac-data.org/
24 Quantization can be done only for onsets, or for onsets and offsets.
25 Threetriplet quarter notes durations sum a half note duration, so the duration of a triplet
quarter note is 2/3 the duration of a standard quarter note.

37
2. BACKGROUND

2 3 4 5 6 7 8 9 10 11

Z . . 6 . . 7 . . 8 . . 9 . . 10 . # .. . # . 11 D
. . . .
2 3 4 5

&\ . . . . . . .
. .
% Z\ D D . D . . . D
Inst 3
Piano

. .
z
z
{z z
z
{z
z 16 z
z
{z 17 z
z
{z
z 18z
z
{z
z
z z
{z
z19 z
z
{z
z20z
z
z
{z
zzz
z
z
{z
zzz z
{
Figure .2.20:
. .Example . . . (top)
. . .of a. piano-roll . . and. # score
... . # representation
. . . . . (bottom). .
12 13 14 15 21

for & . .
an excerpt. of the
. MIDI file . .# .# . .
RWC-MDB-C-2001-27 from Goto
. .RWC
(2003),
database, W.A. Mozart variations on ‘Ah Vous Dirai-je Maman’, K.265/300e.
.
D . D using
%obtained . DLogic. .Pro. 8. D . D . . . D
.
Figs.

z
zz
z. zz{z
z
z z z
{z
z z zz
{z
z z {z z
z z
{z
z z z
z
{z z z
z
{z z
zz
{z
zzz{z
z
zz{z
z
z z
z
{z
z z
2.422 .Supervised
23
. . 24. learning
. 25. . 26 . # ... . # . 27 - 28 29 . . . . .
& . . . .
. . . . . .
. D .of pairs
D consist D
Supervised learning methods attempt to deduce a function from a training
set.%The training data . of. input- data (usually, . vectors), and
desired outputs. After the learning stage, a supervised learning algorithm can
30 z
z
{z
.z zz
{z
z
z of zzz
{z
zzfunction
z
z
z
{z
z. Pfor
.zz
{z
zz
.any. z
z
z
{z
.z. zz.z
{z
z
z. z
z
{32z
z
{z
z
.z
z
z3z
z
z
{z
zz
z
zz. basic
zz z z z
z
{
. Q . the . . . 31Olearning
. .supervised . . . . . .
predict the value the valid input data. The concepts
& . . . .
to understand methods used in Chapter 6 for multiple
f0 estimation .
are briefly described next. . Q.
%
2.4.1 Neural networks
z
zzzz
zz
z
zzz z z
.{zz
z
.z
zz z zzz
{zz z
zzzz.zzz z.z
z
{z
z
zz
zz
zz 35z z
z z
{zzz
zz
z
zzzz z
zz
{z z
z z { z
zz
z
. . . .perceptron . . or .multilayer
. . (MLP) . .network
. . .neural . D
.
33 34 36

& . introduced by Minsky. and Papert (1969). Citing Duda. et .al. (2000),
The multilayer architecture
was first
multilayer neural networks.implement linear discriminants, but in a space where
D E# D E . . D
the% "" mapped.# The "key .power provided by. such
"
inputs have been nonlinearly

37 z
z
.zz
zz
zzz z
zz
{z
zz
z z { z zzz
zz
zzz
z
z z z
z
{z
z
zdata.
z{ z zz
zzzzzzz
z
{z
z
z
z{ zzzz
zzz
z
{zzz
{z
networks is that they admit fairly simple algorithms where the form of the
. Qcan
nonlinearity . . . from
. .be. learned 38
. .training
. . . . . . . . O. P. O . P. . .
39

&
The. Fig. 2.21 shows an example of a MLP. This sample neural network is
. .
composed by three layers: the input layer (3 neurons), a hidden layer (2 neurons)
and% D layer (3. neurons), connected
output D . weighted Dedges.
by .

z
z
zz
z
z
zz
zzzz
z
z
z{z
zzz
zz{ z
z
zz
z
zz
z
zz
zzz
z
{z
z{ z
z
z
zz
z
zzz z
z
z
{z
z
zzz
z{
38
 2.4. SUPERVISED LEARNING

Output values

Output layer

Weigth matrix 2

Hidden layer

Weight matrix 1

Input layer

Input values

Figure 2.21: Multilayer perceptron architecture.

The weights of a multilayer neural network can be learned by the standard

backpropagation algorithm from Rumelhart et al. (1986). Each feature of the
input vector is presented to one neuron of the input layer, whereas each hidden
unit performs the weighted sum of its inputs to form its activation, and each
output unit similarly computes its activation based on the hidden unit signals
(Duda et al., 2000).
A transfer function26 , which is typically a sigmoid, is used to determine
whether a neuron is activated or not according to its input.

Time-delay neural networks

As pointed out by Hush and Horne (1993), time-delay neural networks (Waibel,
1989) are considered as non-recurrent dynamic networks, although essentially
are like static nets traversing temporal series. This kind of nets can model
systems where the output y[t] has a dependence of a limited time interval in the
input u[t]:
 
y(t) = f [u(t − m), ..., u(t), ..., u(t + n)] 2.23 
Using this network, time series can be processed as a collection of static
input-output patterns, related in short-term as a function of the width of the
input window. The TDNN architecture is very similar to a MLP, but the main
difference is that the input layer is also fed with information about adjacent
time frames. Each hidden unit accepts input from a restricted (spatial) range
26 Also called activation function.

39
2. BACKGROUND

Figure 2.22: TDNN architecture. Fig. from Duda et al. (2000).

of positions in the input layer (see Fig. 2.22). A TDNN can be trained with the
same standard backpropagation algorithm used for a MLP.

2.4.2 Nearest neighbors

The nearest neighbors (NN) algorithm is a non-parametric method proposed

by Cover and Hart (1967) to classify objects based on their closest training
examples in a feature space. Given an unknown sample, the NN algorithm
finds in the training set the closest prototype in a n-dimensional feature space,
classifying the test sample in the same class than the prototype.
Unlike neural networks, this is a type of instance-based learning where the
function is only approximated locally, therefore no new classes (or prototypes)
can be directly derived except from the ones that are present in the training
stage. For this reason, the kNN algorithm is sensitive to the local structure of
the data.
The success rates obtained with NN can be improved using a number of k
nearest neighbors. A k nearest neighbor (kNN) algorithm finds the k nearest
neighbors of a sample, assigning it to the class with most representatives through
a voting process.

40
 2.4. SUPERVISED LEARNING

*
* * 2NN
3NN
* * 1NN

4NN

Figure 2.23: Simplified example of NN classification in a two-dimensional feature

space. Using the Euclidean distance, the test sample is classified as the blue
squared class.

Usually, the Euclidean distance is chosen to measure the proximity in

the feature space, though other metrics, such as Manhattan and Minkowsky
distances27 , can be used instead.

27 More information about different metrics used for NN classification can be found in (Duda

et al., 2000), section 4.6.

41
 Music transcription
3
This chapter briefly addresses the human music transcription process, followed
by an analysis of the theoretical issues in automatic transcription from a signal
processing point of view. Finally, the onset detection task is also introduced.
The different metrics used for the evaluation of multiple f0 estimation and onset
detection methods are also discussed.

3.1 Human music transcription

Trained listeners can usually get better transcriptions than automatic systems.
This can be justified since the music transcription problem is similar to speech
recognition, where human training and experience play an important role.
Most automatic transcription systems consider isolated frames (sometimes,
few adjacent frames), whereas humans consider a wider context. Developing
algorithms that consider a wide context is a challenging task due to the high
computational cost required.
Some studies have been done to analyze the processes that trained musicians
use for transcription. It is necessary to have musical skills to write down the
listened notes into a musical notation. This ability is acquired through musical
dictation practice. However, it is not necessary much training for recognizing,
in some way, the chords, the melody and the rhythm in a musical piece, and we
can easily remember and sing a musical piece. Therefore, musical skills can be
split into written skills, related to notation abilities, and aural skills, related to
the perceptual understanding of music.
Memory is very important in the transcription process. Citing Sloboda
(1985), “the way one hears music is crucially dependent upon what one can
remember of past events in the music ... To perceive an event musically (that
is, to recognize at least part of its musical function) is to relate it to past events”.
We can hear key modulations if we can remember the previous key, and a
note or chord has no musical significance without considering the preceding or

43
3. MUSIC TRANSCRIPTION

following events. We can identify relative pitch differences, more than absolute
pitches. Another proof of the importance of the musical context is that pitches
can be very hard to identify in a confusing context like, for instance, when two
different songs are heard simultaneously.
Klapuri et al. (2000) performed a test to measure the pitch identification
ability of trained musicians using isolated chords. The results were compared
with those obtained using an automatic transcription system, and only the
two most skilled subjects performed better than the computational approach,
showing that it is not easy to analyze notes out of context.
Hainsworth (2003) proposed a test where trained musicians where asked to
answer how did they perform transcription. A common pattern was found. The
first step was to do a rough structural analysis of the piece, breaking the song
into sections, finding repetitions, and in some cases marking key phrases. Then,
a chord scheme or the bass line were detected, followed by the melody. Finally,
the inner harmonies were heard by repeated listening, building up a mental
representation. According to Hainsworth (2003), “no-one transcribes anything
but simple music in a single pass”.
The auditory scene analysis (ASA) is a term proposed by Bregman (1990)
to describe the process by which the human auditory system organizes sound
into perceptually meaningful elements. In computational analysis, the related
concept is called computational auditory scene analysis (CASA), which is closely
related to source separation and blind signal separation. The key aspects of
the ASA model are segmentation, integration, and segregation. The grouping
principles of ASA can be categorized into sequential grouping cues (those that
operate across time, or segregated) and simultaneous grouping cues (those
that operate across frequency, or integrated). In addition, schemas (learned
patterns) play an important role. Mathematical formalizations to the field of
computational auditory perception have also been proposed, for instance, by
Smaragdis (2001) and Cont (2008).
The main advantage for humans when transcribing music is our unique
capability to identify patterns and our memory, which allows us to predict
future events. Using memory in computational transcription usually implies a
huge computational cost. It is not a problem to include short term memory , but
finding long term repetitions means keeping alive many hypothesis for various
frames, which is a very costly task. Solving certain ambiguities that humans
can do using long-term memory remains as a challenge. An excellent analysis of
prediction, expectation and anticipation of musical events from a psychological
point of view is done by Cont (2008), who also proposes computational
anticipatory models to address several aspects of musical anticipation. Symbolic
music sequencies can also be modeled and predicted in some degree (Paiement
et al., 2008), as they are typically composed of repetitive patterns.

44
 3.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

Within a short context, a trained musician can not identify any musical
information when listening to a 50 ms isolated segment of music. With a 100
ms long segment, some rough pitch estimation can be done, but it is still difficult
to identify the instrument. Using longer windows, the timbre becomes apparent.
However, multiple f0 estimation systems that perform the STFT can estimate
the pitches using short frames1 . Therefore, probably computers can do a better
estimate in isolated time frames, but humans can transcribe better within a
wider context.

3.2 Multiple fundamental frequency estimation

Single f0 estimation methods attempt to find the fundamental frequency in
signals with at most one harmonic source sounding at each time. A single
harmonic signal can be expressed as the sum of the harmonic part and the
residual z:

H
X  
x[n] ≈ Ah cos(hω0 n + φh ) + z[n] 3.1 
h=1

where ω0 = 2πf0 . The relation in Eq. 3.1 is approximated for practical use, as
the signal can have some harmonic deviations.
A multiple f0 estimation method assumes that there can be more than one
harmonic source in the input signal. Formally, the sum of M harmonic sources
can be expressed as:

Hm
M X
X  
x[n] ≈ Am,hm cos(hm ωm n + φm [n]) + z̄[n] 3.2 
m=1 hm =1

where z̄[n] is the sum of the residuals of the sources.

It is generally admitted that most single f0 methods are not adequate
for multiple f0 estimation. A study of the performance of different single f0
estimation functions applied to the analysis of polyphonic signals can be found
in (Yeh, 2008), p. 19.
The problem is that the polyphonic issue is much more complex than the
single f0 estimation. First, the number of sources M must be inferred, in
contrast to the single f0 estimation, where the only decision is to determine
whether there is sound or silence. Estimating the number of sounds, or
polyphony inference, is a challenging task itself. The noise model is also more
complex than in single f0 signals. In real polyphonic music, besides transients,
there can be unpitched sounds (like drums), which usually have a short duration.
1 The typical frame length used to detect multiple f0 when using the STFT is about 93 ms.

45
3. MUSIC TRANSCRIPTION

Noise suppression techniques have been proposed in the literature2 to allow the
subtraction of additive noise from the mixture. And the third major issue is
that, besides the source and noise models, in multiple f0 estimation there is a
third model, which is probably the most complex: the interaction between the
sources.
For instance, consider two notes playing simultaneously within an octave
interval. As their spectrum shows the same harmonic locations than the lowest
note playing alone, some other information (such as the energy expected in
each harmonic for a particular instrument) is needed to infer the presence of
two notes. This issue is usually called octave ambiguity.
According to Klapuri (2004), in contrast to speech, the pitch range is wide3
in music, and the sounds produced by different musical instruments vary a lot in
their spectral content. The harmonic pattern of an instrument is also different
from low to high notes. And transients and the interference of unpitched content
in real music has to be addressed too.
On the other hand, in music the f0 values are temporally more stable than
in speech. Citing Klapuri (2004), it is more difficult to track the f0 of four
simultaneous speakers than to perform music transcription of four-voice vocal
music.
As previously discussed in Sec. 2.3.1, consonant intervals are more frequent
than dissonant ones in western music. Therefore, pleasant chords include
harmonic components of different sounds which coincide in frequency (harmonic
overlaps). Harmonic overlaps and beating are the main effects produced by the
interaction model, and they are described below.

3.2.1 Harmonic overlap

As pointed out by Klapuri (1998), two sources with fundamental frequencies fa
and fb are harmonically related when fa = m n fb , being m and n positive integer
numbers. In this case, every nth harmonic of the source a overlaps every mth of
the source b. As mentioned above, this scenario is very frequent in music.
When two sounds are superposed, the corresponding wave functions are
summed. When there is a harmonic overlap, two simple harmonic motions
with the same frequency, but different amplitudes and phases are added. This
produces another simple harmonic motion with the same frequency but different
amplitude and phase. Therefore, when two harmonics are overlapped, two
sinusoids of the same frequency are summed in the waveform, resulting a signal
with the same frequency and which magnitude depends on their phase difference.
Considering K sinusoids overlap, the resulting sinusoid can be written as:
2 For a review of noise estimation and suppression methods, see (Yeh, 2008), chapter 4.
3 The tessitura of an instrument is the f0 range that it can achieve.

46
 3.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

K
X  
A cos(ωn + φ) = Ak cos(ωn + φk ) 3.3 
k=1

Using trigonometric identity, the resulting amplitude (Yeh and Roebel, 2009)
can be calculated as:
v
#2 " K #2
 
u" K
u X X
A= t Ak cos(φk ) + Ak sin(φk ) 3.4 
k=1 k=1

From which the estimated amplitude A of two overlapping partials with the
same frequency, different amplitude, and phase difference φ∆ is:
q  
A = A21 + A22 + 2A1 A2 cos(φ∆ ) 3.5 
As pointed out by Yeh and Roebel (2009), two assumptions are usually
made for analyzing overlapping partials: the additivity of linear spectrum and
the additivity of power spectrum. The additivity of linear spectrum A = A1 +A2
assume that the two sinusoids
p are in phase, i.e., cos(φ∆ ) = 1. The additivity of
power spectrum A = A21 + A22 is true when cos(φ∆ ) = 0.
According to Klapuri (2003a), if one of the partial amplitudes is significantly
greater than the other, as is usually the case, A approaches the maximum of
the two. Looking at Eq. 3.5, this assumption is closely related to the additivity
of power spectrum, which experimentally (see Yeh and Roebel, 2009) obtains
better amplitude estimates than considering cos(φ∆ ) = 1.
Recently, Yeh and Roebel (2009) proposed an expected overlap model to get
a better estimation of the amplitude when two partials overlap, assuming that
the phase difference is uniformly distributed.

3.2.2 Beating
If two harmonics of different sources do not coincide in frequency4 , but they
have similar amplitude and small frequency difference, interference beats can
be perceived (see Fig. 3.1).
As pointed out by Wood (2008), p. 158, the physical explanation of
dissonance is that we hear unpleasant beats. Beats are periodic variations of
loudness, and the frequency of the beats depend on the frequency difference of
the two tones.
Even when the frequency difference between two partials is not small enough
to produce a perceptive beating, some amount of beating is always produced
between a pair of harmonics, even when they belong to the same source. The
4 For instance, due to slight harmonic deviations.

47
3. MUSIC TRANSCRIPTION

s1 (t)

s2 (t)

s(t) = s1 (t) + s2 (t)

Figure 3.1: Interference tones of two sinusoidal signals of close frequencies. Fig.
extracted from http://www.phys.unsw.edu.au/jw/beats.html

beating effect generates spectral components not belonging to any original

source (see Fig. 3.2), producing ghost fundamental frequencies, and it also alters
the original partial amplitudes in the spectrum.

3.2.3 Evaluation metrics

There are a number of evaluation issues for multiple f0 estimation. First, it is
difficult to get real data aligned with the ground-truth pitches, and doing this
alignment by hand is a time-consuming task. A simple database can be built
synthesizing data from MIDI files, but synthesized data is not equivalent to real
sounds.
One of the reason for this difference is reverberation. Reverberation, or
reverb, is created when a sound is produced in a closed space, causing a large
number of echoes to build up and then slowly decaying as the sound is absorbed
by the walls and air (Lloyd, 1970). Reverberation overlaps preceding sounds
with following ones. As pointed out by Beauchamp et al. (1993) and Yeh
et al. (2006), when a monophonic recording is carried out in a reverberant
environment, the recorded signal can become polyphonic. Citing Yeh (2008),
the reverberated parts are quite non-stationary, increasing the complexity of the
analysis. Besides reverb, other musical production effects like chorus or echo,
which are common in commercial recordings, can complicate the analysis.
Other difference is that synthesized signals from MIDI files typically have
constant envelopes, making unnatural sounds for instruments with varying
envelopes like, for instance, a sax or a violin. Real sounds are usually much
more variable and less static than synthesized sounds.

48
 3.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

45
G3 "dosolsplitnohwr.txt" using ($0*(22100/2048)):1

40

35

30 C3

25

20

15

10

0
0 200 400 600 800 1000 1200 1400 freq (Hz)

beating
frequency

Figure 3.2: Example spectrum of two piano sounds with fundamental

frequencies C3 (130.81 Hz) and G3 (196 Hz). A beating component appears
at frequency 65 Hz, corresponding to a C2 ghost pitch.

Different metrics have been used in the literature to evaluate polyphonic

estimation methods. They can be classified into frame-by-frame metrics
(fundamental frequencies are evaluated within single frames), and note-based
metrics (note onsets and durations are also taken into account). The former
are used to evaluate most multiple f0 estimation methods, whereas note-based
metrics are suitable for the evaluation of those approaches that also perform f0
tracking5 .

Frame-based evaluation
Within a frame level, a false positive (FP) is a detected pitch which is not
present in the signal, and a false negative (FN) is a missing pitch. Correctly
detected pitches (OK) are those estimated pitches that are also present in the
ground-truth.
A commonly used metric for frame-based evaluation is the accuracy, which
can be defined as:

ΣOK  
Acc = 3.6 
ΣOK + ΣF P + ΣF N
5 Thisterm refers to the tracking of the f0 estimates along consecutive frames in order to
add a temporal continuity to the detection.

49
3. MUSIC TRANSCRIPTION

Alternatively, the error metric can be defined in precision/recall terms. The

precision is the percentage of the detected pitches that are correct, whereas the
recall is the percentage of the true pitches that were found with respect to the
actual pitches.

ΣOK  
Prec = 3.7 
ΣOK + ΣF P

ΣOK  
Rec = 3.8 
ΣOK + ΣF N
The balance between precision and recall, or F-measure, which is commonly
used in string comparison, can be computed as:

2 · Prec · Rec ΣOK  

F-measure = = 3.9 
Prec + Rec ΣOK + 12 ΣF P + 12 ΣF N
Precision, recall and F-measure can also be expressed in percentages,
multiplying by 100 the expressions above. Note that F-measure yields higher
values than accuracy having the same amount of errors.
An alternative metric based on the “speaker diarization error score” from
the NIST6 was proposed by Poliner and Ellis (2007a) to evaluate frame-level
polyphonic transcriptions. The NIST metric consists of a single error score
which takes into account substitution errors (mislabeling an active voice, Esubs ),
miss errors (when a voice is truly active but results in no transcript, Emiss ),
and false alarm errors (when an active voice is reported without any underlying
source, Ef a ).
This metric was proposed to avoid counting errors twice as accuracy of F-
measure do in some situations. For instance, using accuracy, if there is a C3
pitch in the reference ground-truth but the system reports a C4 , two errors
(a false positive and a false negative) are counted. However, if no pitch was
detected, only one error would be reported.
To compute the total error (Etot ) in T frames, the estimated pitches at every
frame are denoted as Nsys , the ground-truth pitches as Nref , and the number
of correctly detected pitches as Ncorr , which is the intersection between Nsys
and Nref .
PT
max(Nref (t), Nsys (t)) − Ncorr (t)  
t=1
Etot = PT 3.10 
t=1 Nref (t)

The substitution, miss and false alarm errors are defined as follows:
PT
min(Nref (t), Nsys (t)) − Ncorr (t)  
t=1
Esubs = PT 3.11
 
t=1 Nref (t)
6 National Institute of Standards and Technology.

50
 3.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

PT
max(0, Nref (t) − Nsys (t))  
t=1
Emiss = PT 3.12 
t=1 Nref (t)
PT
max(0, Nsys (t) − Nref (t))  
t=1
Ef a = PT 3.13 
t=1 Nref (t)

Poliner and Ellis (2007a) suggest that, as in the universal practice in the
speech recognition community, this is probably the most adequate measure,
since it gives a direct feel for the quantity of errors that will occur as a proportion
of the total quantity of notes present.
To summarize, three alternative metrics are used in the literature to evaluate
multiple f0 estimation systems within a frame level: accuracy (Eq. 3.6), F-
measure (Eq. 3.9), and total error (Eq. 3.10).
The accuracy is the most widely used metric for frame by frame evaluation.
The main reason for using accuracy instead of F-measure is that an equilibrate
balance between precision and recall is probably less adequate for this task.
Typically, multiple f0 estimation methods obtain higher precision than recall.
This occurs because some analyzed mixtures contain many pitches with
overlapped harmonics that can be masked by the other components. An
experiment was carried out by Huron (1989) to study the limitations in
listeners abilities to identify the number of concurrent sounding voices. The
most frequent type of confusion was the underestimation of the number of
sounds. Some pitches can be present in the signal, but they can become almost
unhearable, and they are also very difficult to detect analytically. For instance,
when trying to listen an isolated 93 ms frame with 6 simultaneous pitches, we
usually tend to underestimate the number of sources.

Note-based evaluation

Instead of counting the errors at each frame and summing the result for all
the frames, alternative metrics have been proposed to evaluate the temporal
continuity of the estimate. Precision, recall, F-measure and accuracy are also
frequently used for note level evaluation. However, it is not trivial to define
what is a correctly detected note, a false positive, and a false negative.
The note-based metric proposed by Ryynänen and Klapuri (2005) considers
that a reference note is correctly transcribed when their pitches are equal,
the absolute difference between their onset times is smaller than a given
onset interval, and the transcribed note is not already associated with another
reference note. Results are reported using precision, recall, and the mean overlap
ratio, which measures the degree of temporal overlap between the reference and
transcribed notes.

51
3. MUSIC TRANSCRIPTION

The multiple f0 tracking evaluation metric used in MIREX (2007, 2008)

reports the results in two different ways. In the first setup, a returned note is
assumed correct if its onset is within ±50 ms of a reference note and its f0 is
correct, ignoring the offset values. In the second setup, on top of the above
requirements, a correct note was required to have an offset value within 20% of
the reference note duration around the reference offset, or within 50 ms.
Daniel et al. (2008) introduced a perceptual error metric for multiple f0
evaluation. To estimate the perceptual relevance of note insertions, deletions,
replacement, note doubling, etc., thirty-seven subjects were asked to obtain
a subjective scale of discomfort for typical errors. Then, a set of weighting
coefficients related to the errors made when detecting a wrong octave, fifth,
other intervals, deletion, duration, and onset, were estimated to get a perceptive
F-measure. In this metric, octave errors have a lower impact than fourth errors,
for instance.
The perceptual measure proposed by Fonseca and Ferreira (2009) for note-
based evaluation considers different metrics for decay (percussive) and sustained
sounds. The decay-based detection only takes into account onset and pitch,
whereas the sustained sounds are evaluated by overlapping the original and
transcribed piano-rolls with a tolerance degree (i.e., considering pitch, onset
and offset). As the nature of the sounds is unknown a-priori, the final score is
set as the average of the decay and sustain scores.

3.3 Onset detection

Onset detection refers to the detection of the beginnings of discrete events in
audio signals. These events can be either pitched or unpitched sounds. It is a
component of the segmentation process which aims to divide a musical signal
into smaller units (individual notes, chords, drum sounds, etc.).
This task can be useful for a number of applications. Some tempo estimation
and beat tracking methods use the onsets information. Onset times are also
useful for multiple f0 estimation and source separation tasks, as the beginning
of notes can be used for segmentation. An onset list can also serve as a robust
fingerprint for a song, to identify it in a database. Some music classification
approaches, like (Lidy et al., 2007), use onsets as input features. Other
applications include music editing, as they can be used to divide the song into
logical parts, and audio/video synchronization.
Some works define onset as the time instant when the attack interval or
when a musical sound begins, whereas others, like (Bello et al., 2005), consider
onsets as the time instants when a transient starts.
Onsets can be categorized according to the source which produce them. For
instance, Tan et al. (2009) classifies onsets as those produced by unpitched

52
 3.3. ONSET DETECTION

t (sec)

Figure 3.3: Example of a guitar sound waveform. The actual onsets are marked
with dashed vertical lines.

sounds like drums, pitched percussive onsets like pianos, and pitched non-
percussive onsets like bowed strings. Unpitched and pitched percussive sounds
produce hard onsets, whereas pitched non-percussive timbres usually generate
soft onsets.

3.3.1 Evaluation metrics

Rigorous evaluation of onset detection is a complex task (Rodet et al., 2004).
The evaluation results of onset detection algorithms are in some cases not
comparable, as they depend much on the database used for the experiments.
According to Moore (1997), the human ear can not distingish between two
transients less than 10 ms apart. However, as Bello et al. (2005) points, correct
matches in the evaluation usually imply that the target and detected onsets are
within a 50 ms window, to consider the inaccuracy of the hand labelling process.
In the literature, there is agreement to express the onset detection results in
precision/recall and F-measure terms, similarly to Eqs. 3.7, 3.8, and 3.9 used
for multiple f0 estimation. A false positive (FP) is a detected onset that was
not present in the signal, and a false negative (FN) is an undetected onset.
The precision is the ratio of the correctly detected onsets, and the recall is the
ratio between the true onsets that were found with respect to the ground-truth
onsets.

53
 State of the art
4
This chapter presents an overview of previous studies for single f0 estimation,
followed by a deeper review and discussion of different multiple f0 estimation
and onset detection methods.

4.1 Single fundamental frequency estimation

A number of single fundamental frequency estimation systems have been
developed in the time and the frequency domains. For a review of different
methods, see (Rabiner et al., 1976), (Hess, 1983), and (Gerhard, 2003). Most of
the proposed approaches were initially developed for speech signals1 , but they
have also been successfully applied to musical harmonic sounds.

4.1.1 Time domain methods

Time domain methods look for a repetitive pattern in the signal, corresponding
to the fundamental period. A widely used technique is the autocorrelation
function, which is defined for a signal x[t] with a frame length W as:

X−1
t+W  
ACFx [τ ] = x[k]x[k + τ ] 4.1 
k=t

where τ represents the lag value. The peaks of this function correspond to
multiples of the fundamental period. Usually, autocorrelation methods select
the highest non-zero lag peak over a given threshold within a range of lags.
However, this technique is sensitive to formant structures, producing octave
errors. As Hess (1983) points out, some methods like center clipping (Dubnowski
1 Detecting the fundamental frequency in speech signals is useful, for instance, for prosody

analysis. Prosody refers to the rhythm, stress, and intonation of connected speech.

55
4. STATE OF THE ART

et al., 1976), or spectral flattening (Sondhi, 1968) can be used to attenuate these
effects.
The squared difference function (SDF) is a similar approach to measure
dissimilarities, and it has been used by de Cheveigné and Kawahara (2002) for
the YIN algorithm.

X−1
t+W
2
 
SDFx [τ ] = (x[k] − x[k + τ ]) 4.2 
k=t

The YIN method computes the cumulative mean normalized difference

function (SDF0 ), which is obtained by dividing each value of the SDF by its
average over shorter-lag values.

1 if τ = 0,  
SDF0x [τ ] = SDF x [τ ] 4.3 

(1/τ ) τj=1 SDFx [j]
P otherwise

The main advantage of using the SDF0 function is that it tends to remain
large at low lags, dropping below 1 only where SDF falls below the average.
Basically, it removes dips and lags near zero avoiding super-harmonic errors,
and normalization makes the function independent of the absolute signal level.
An absolute threshold is set, choosing the first local minimum of SDF0 below
that threshold. If none is found, the global minimum is chosen instead. Once
the lag value τ is selected, a parabolic interpolation of immediate neighbors is
done to increase the accuracy of the estimate, obtaining τ 0 , and the detected
fundamental frequency is finally set as f0 = fs /τ 0 .
YIN is a robust and reliable algorithm that have been successfully used as
basis for singing voice transcription methods, like the one proposed by Ryynänen
and Klapuri (2004).

4.1.2 Frequency domain methods

Usually, methods in the frequency domain analyze the locations or the distance
between hypothetical partials in the spectrum.

Cepstrum

The real cepstrum of a signal is the inverse Fourier transform of the logarithm
of the magnitude spectrum.
 
CEPx [τ ] = IDFT{log(|DFT(x[n])|} 4.4 
It was introduced for fundamental frequency estimation of speech signals by
Noll and Schroeder (1964), who gave a complete methodology in (Noll, 1967).

56
 4.1. SINGLE FUNDAMENTAL FREQUENCY ESTIMATION

Since temporal periodicities in a speech signal cause periodic ripples in the

amplitude spectrum, the cepstrum gives the frequency of the ripple, which is
inversely proportional to the fundamental frequency of the speech.
ACF-based fundamental frequency estimators have some similarities with
cepstrum-based methods. The ACF of a time-domain signal x[n] can also be
expressed using the DFT :

K−1 
1 X

2πτ k
   
2 2
ACFx [τ ] = IDFT{|DFT(x[n])| } = cos |X[k]| 4.5 
K K
k=0

Note that the cosine factor emphasizes the partial amplitudes at those
harmonic positions multiple of τ . The main difference between autocorrelation
and cepstrum is that autocorrelation uses the square of the DFT, and the
cepstrum performs the logarithm. Squaring the DFT causes to raise spectral
peaks but also the noise. Using the logarithm flats the spectrum, reducing noise
but also the harmonic amplitudes.
Therefore, as pointed out by Rabiner et al. (1976), the cepstrum performs
a dynamic compression over the spectrum, flattening unwanted components
and increasing the robustness for formants, but rising the noise level, whereas
autocorrelation emphasizes spectral peaks in relation to noise, but raising the
strength of spurious components.
Both ACF and cepstrum-based methods can be classified as spectral location
f0 estimators.

Spectral autocorrelation

The main drawback of the spectral location f0 estimators is that they are
very sensitive to harmonic deviations from their ideal position. Some methods,
like the one proposed by Lahat et al. (1987), perform autocorrelation over the
spectrum.

2
K/2−τ −1
X  
ACFSX [τ ] = |X[k]||X[k + τ ]| 4.6 
K
k=0

The maximum ACFS value is usually found when τ = f0 (K/fs ). Spectral

autocorrelation is more robust against inharmonicity than ACF, as the spectrum
can be shifted without affecting the detection results.
Methods using spectral autocorrelation can be classified as a spectral interval
f0 estimators, as they look for harmonic intervals rather than their locations.

57
4. STATE OF THE ART

fmin

f0
fmax

Figure 4.1: Maximum likelihood from Noll (1969).

Pattern matching in the frequency domain

Other approaches compare the measured spectrum with a given harmonic

spectral pattern. The spectral pattern can be either a specific spectral model
or a sequence of equally spaced components, which is often called a harmonic
comb. A comb filter is composed by a set of equally spaced pass-bands. In
the case of the optimum comb filter algorithm, the location of the passbands
coincide with the harmonic locations.
Brown (1992) uses a constant-Q transform in a preprocessing stage to
correlate the spectrum with an ideal harmonic pattern, which consists of ones
at the positions of harmonic frequency components. The f0 is estimated by
looking at the position of the best approximation to the ideal pattern, which
corresponds to the greatest cross-correlation peak.
The Maximum likelihood2 (ML) algorithm proposed by Noll (1969) searches
though a set of possible ideal spectra and chooses the one which best matches
the shape of the input spectrum (see Fig. 4.1). The system is suitable for tuned
pitched sounds, as the base frequencies of ideal spectra coincide with the musical
pitches.
The method proposed by Doval and Rodet (1993) can be also placed in the
harmonic matching class, as it constructs a histogram on the selected intervals
by computing the value of the likelihood function for each interval.
An interesting pattern matching approach is the two-way mismatch (TWM)
algorithm from Maher (1990); Maher and Beauchamp (1994). It is based on
the comparison between the partials obtained from the STFT and predicted
sequences of harmonics relative to the f0 . The differences between the measured
and predicted partials are referred as the mismatch error, which is calculated in
two ways, as shown in Fig. 4.2. The first measures the frequency difference
2 Citing (Davy, 2006a), the likelihood can be seen as a similarity measure between the

signal and the model, via its parameters. The expectation maximization (EM) algorithm by
Moon (1996) can be used for maximum likelihood parameter estimation.

58
 4.1. SINGLE FUNDAMENTAL FREQUENCY ESTIMATION

Figure 4.2: Two way mismatch procedure from Maher and Beauchamp (1994).

between each measured partial and its nearest harmonic neighbor in the
predicted sequence, whereas the second measures the mismatch between each
predicted harmonic and its nearest neighbor in the measured sequence. Each
match is weighted by the amplitudes of the observed peaks. This method tries to
reduce octave errors, applying a penalty to missing and extra harmonics relative
to the predicted pattern. The methodology was also used for duet3 separation4 .
Cano (1998) introduced some modifications over the TWM to improve the
original SMS analysis developed by Serra (1997). These modifications include
a pitch dependent analysis window using adaptive window length, a more
restrictive selection of spectral peaks to be considered, f0 tracking using short-
term history to choose between candidates with similar TWM error, to restrict
the frequency range of possible candidates, and to discriminate between pitched
and unpitched parts.

4.1.3 Perceptual models

Some f0 estimation methods use perceptual models of human hearing. The
unitary model of the auditory system, introduced by Meddis and Hewitt
(1991a,b), can estimate the f0 of a signal by measuring the periodicity of

3A duet is composed by two melodies playing simultaneously.

4 Forthis reason, this algorithm could also be considered as a multiple f0 estimation system
belonging to the joint estimation methods category (see Sec. 4.2.3).

59
4. STATE OF THE ART

the time-domain amplitude envelope. This model represent a tradeoff between

spectral location and spectral interval methods.
As pointed out by Klapuri (2003a), time domain periodicity analysis
methods are prone to errors in f0 halving, whereas frequency domain methods
are prone to errors in f0 doubling. This is because the time-domain signal is
periodic at half the f0 (twice the fundamental period), whereas the spectrum
is periodic at double the f0 rate. The unitary model gets a good compromise
between both.
As pointed out by Klapuri and Astola (2002), the unitary model is widely
accepted as a psychoacoustically valid mid-level representation. To compute
it (see (Meddis and Hewitt, 1991a) and (Duda et al., 1990) for details),
a cochlear frequency analysis is first done using an auditory filter-bank.
Then, a simulation of the hair cells5 analysis is performed through half-wave
rectification6 , compression and low-pass filtering of the signals at each frequency
channel. The periodicity at each channel is estimated through the ACF function,
yielding a correlogram, which is a three-dimensional representation of time,
frequency and ACF lag. Finally, the summary autocorrelation function (SACF),
is computed, summing the ACF across channels. The greatest value of the SACF
function is used as an indicator of the perceived pitch.
Computationally efficient implementations of the SACF have been proposed
by Tolonen and Karjalainen (2000), and Klapuri and Astola (2002). For a
better coverage about pitch perception models, see the review from de Cheveigné
(2005).

4.1.4 Probabilistic models

As pointed out by Roads (1996), a small time window is often not enough for a
human to identify pitch, but when many frames are played one after another, a
sensation of pitch becomes apparent. This is the main motivation to introduce
probabilistic models for f0 tracking, which can be used to refine the f0 estimate.
Intuitively, a simple f0 tracking approach would consist in giving preference
to f0 hypotheses that are close to the hypothesis of the last time frame. A
more reliable method is to use statistical models, like hidden Markov models
(HMMs), which track variables through time. HMMs are state machines, with
a hypothesis available for the output variable at each state. At each time frame,
the HMM moves from the current state to the most likely next state, based on
the input to the model and the state history which is represented in the current
state.
5 Haircells convert cochlear movements into neural impulses.
6 The half-wave rectification of a time-domain signal keeps the positive values, zeroing the
negative ones.

60
 4.1. SINGLE FUNDAMENTAL FREQUENCY ESTIMATION

Figure 4.3: Combinations of note models and the musicological model from
Ryynänen and Klapuri (2004).

A HMM consists of a number of states, the state transition probabilities, the

observation likelihood distributions and the initial and final state probabilities.
The state transition probabilities of a HMMs are learned in an unsupervised
way using the Baum-Welch algorithm (see (Rabiner, 1989) for details). Once
the parameters have been estimated, the state sequence that maximizes the
probability of the observed data can be efficiently computed using the Viterbi
algorithm, described by Viterbi (1967).
HMMs have been used by Ryynänen and Klapuri (2004) for a singing voice
transcription system. First, the fundamental frequency, voicing, accent and
metrical accent are extracted from the audio signal. The fundamental frequency
and voicing are estimated using the YIN algorithm (voicing corresponds to
SDF0x , see Eq. 4.3), and the f0 is rounded to that of its nearest MIDI pitch.
Then, two probabilistic models are used: a note event model and a musicological
model. Note events are described using a HMM for each pitch. The HMMs are
composed of three states (which can be intuitively identified with the attack,
sustain and silence stages), and their inputs are the extracted features (pitch
difference7 , voicing, accent and metrical accent). The musicological model
weights transitions between notes using note n-gram probabilities which are
dependent on the key of the song, which is estimated using the algorithm
proposed by Viitaniemi et al. (2003).

7 The frequency difference in semitones between the measured f and the nominal pitch of
0
the modeled note.

61
4. STATE OF THE ART

Finally, the two models are combined into a network (see Fig. 4.3), and
the most probable path is found according to the likelihoods given by the note
models and the musicological model. The system obtained half amount of errors
than the simple f0 estimation rounded to MIDI pitch, proving the capability of
probabilistic models for this task.

4.2 Multiple fundamental frequency estimation

Since the first polyphonic transcription attempt from Moorer (1975) to tran-
scribe duets with some interval constraints, many different techniques have been
proposed for multiple f0 estimation. As pointed out by Klapuri (2004), it is
difficult to classify these methods using a single taxonomy, as they are very
complex and usually combine several processing principles.
For instance, multiple f0 estimation systems can be categorized according
to their mid-level representation (time-domain, STFT, wavelets, auditory filter
banks, etc.), but also to their scope (some methods need a-priori information
about the instrument to be transcribed, whereas others can be used to analyze
generic harmonic sounds), to their capability for modeling varying timbres
(for instance, statistical parametric methods can model varying time-frequency
envelopes like those of sax sounds, whereas non-parametric methods can
only handle fixed spectral patterns, like piano sounds), or by the way they
can estimate the interactions between sources (iterative and joint estimation
methods).
In this work, the proposed categorization is based on the core methodology,
rather than the mid-level representation used in the single f0 estimation
taxonomy. Existing approaches have been classified into salience functions,
iterative cancellation, joint estimation, supervised learning, unsupervised learn-
ing, matching pursuit, Bayesian models, statistical spectral models, blackboard
systems, and database matching methods. An analysis of the strengths and
limitations for each of these categories is done in Sec. 4.3.

4.2.1 Salience methods

Salience methods try to emphasize the underlying fundamental frequencies by
applying signal processing transformations to the input signal. In these ap-
proaches the core of the detection process relies on the mid-level representation.
Tolonen and Karjalainen (2000) perform the SACF over the signal using
an auditory filter bank. The SACF is processed to remove the near-zero
lag components and the peaks that are multiple of the other peaks found.
The resulting function is called enhanced summary autocorrelation function
(ESACF).

62
 4.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

Figure 4.4: Iterative cancellation method from Klapuri (2003a).

The method from Peeters (2006) combines a temporal representation (time-

domain ACF and real cepstrum) with a spectral representation (spectral
autocorrelation) to reduce octave ambiguities. The best results were reported
when combining the spectral autocorrelation function with the cepstrum.
Zhou et al. (2009) propose an efficient method which relies on a novel time-
frequency representation called resonator time-frequency image (RTFI). The
RTFI (Zhou and Mattavelli, 2007) selects a first order complex resonator filter
bank to implement a frequency dependent time-frequency analysis. Harmonic
components are extracted by transforming the RTFI average energy spectrum
into a relative energy spectrum. Then, a preliminary estimation of pitch
candidates is done by converting the RTFI average spectrum into a pitch energy
spectrum (PES) and a relative pitch energy spectrum (RPES). The information
about harmonic components and pitch candidates are combined to remove extra
pitches. Finally, the remaining candidates are filtered out by using a smoothness
criterion to remove extra pitches again, considering only cases for which the
frequency ratio of two candidates is 2, 3 or 4.

4.2.2 Iterative cancellation methods

Some methods estimate the most prominent f0 , subtracting it from the mixture
and repeating the process for the residual signal until a termination criterion.
In the method proposed by Klapuri (2003a), Fig. 4.4, the spectrum of the
signal is warped on a logarithmic frequency scale to compress the spectral
magnitudes and remove the noise. The processed spectrum is analyzed into
a 2/3 octave filter bank, and f0 weights are computed for each band according
to the normalized sum of their partial amplitudes. The results are combined by
summing the squared band-wise weights, taking inharmonicity (Eq. 2.17) into
account. The spectral components of the fundamental frequencies that have the
highest global weights are smoothed using the algorithm described in (Klapuri,

63
4. STATE OF THE ART

Figure 4.5: Probabilistic framework from Ryynänen and Klapuri (2005).

2001) before subtracting them from the mixture. The weights of each candidate
are calculated again after smoothing, and the highest recalculated global weight
determines the resulting f0 . The process stops when the maximum weight
related to the signal-to-noise ratio (SNR) is below a fixed threshold.
Klapuri (2005) proposed an alternative method using an auditory filter
bank. The signal at each subband is compressed, half-wave rectified and low-
pass filtered. Then, similarly to the SACF, the results are combined across
channels, but in this method magnitude spectra are summed across channels
to obtain a summary spectrum. The most salient f0 is computed using an
approximated 1/h spectral envelope model8 , to remove the source for the
mixture while keeping in the residual most of the energy of higher partials.
This model is similar to (Klapuri, 2006b), where the input signal is flattened
(whitened) to reduce timbral-dependant information, and the salience for each
f0 candidate is computed as a 1/h weighted sum of its partials. This same partial
weighting scheme is performed by Klapuri (2008) using a computationally
efficient auditory model.
The system introduced by Ryynänen and Klapuri (2005) embeds the multiple
f0 estimator from Klapuri (2005) into a probabilistic framework (see Fig.
4.5), similarly to the method from Ryynänen and Klapuri (2004) for singing
transcription. As in the latter work, a note event model and a musicological
model, plus a silence model are used, and note events are described using a
HMM for each pitch. The HMM inputs are the pitch difference between the
measured f0 and the nominal pitch of the modeled note, the f0 salience, and the
onset strength (positive changes in the estimated strengths of f0 values). The
musicological model controls transitions between note HMMs and the silence
model using note bigram probabilities which are dependent on the estimated
8 Spectral amplitudes decrease according to the partial number h, like in a sawtooth signal.

64
 4.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

Figure 4.6: Overview of the joint estimation method from Yeh (2008).

key. Like in (Ryynänen and Klapuri, 2004), the acoustic and musicological
models are combined into a network which optimal path is found using the
token-passing algorithm from Young et al. (1989).
Other examples of iterative cancellation methods are those proposed by Wan
et al. (2005), Yin et al. (2005), and Cao et al. (2007).

4.2.3 Joint estimation methods

These methods evaluate a set of possible hypotheses, consisting of f0 combina-
tions, to select the best one without corrupting the residual at each iteration.
Time-domain methods for joint cancellation of multiple f0 hypothesis have
been proposed by de Cheveigné (1993, 2005). The hypotheses are cancelled using
a cascade of filters, and the combination selected is the one that minimizes the
residual. In the experiments done by de Cheveigné (2005), different iterative
cancellation methods are compared with the joint approach, showing that joint
cancellation outperforms the iterative cancellation results.
The method proposed by Yeh (2008) evaluates a set of multiple f0 hypotheses
without cancellation. An adaptive noise level estimation (see Fig. 4.6) is first
done using the algorithm described in (Yeh and Roebel, 2006), in order to
extract only the sinusoidal components. Then, given a known number of sources,
the fundamental frequencies are obtained using the method described in (Yeh
et al., 2005). At each spectral frame, to reduce the computational cost, a set of
f0 candidates are selected from the spectral peaks using a harmonic matching
technique9 . Each f0 hypothesis is related to a hypothetical partial sequence
(HPS). The HPS is a source model with estimated frequencies and amplitudes
obtained by partial selection and overlapping partial treatment. Partials are
identified with spectral peaks within a tolerance deviation from their ideal
position. In case that more than one peak is in the tolerance range, the peak
forming a smoother HPS envelope is selected. The amplitudes of overlapped
partials in the combination are estimated by using linear interpolation, similarly
to (Maher, 1990), and a set of rules. Finally, HPS are flattened by exponential
compression.
Once HPS are estimated, a score function for a given hypothesis is
calculated taking into account, for each hypothetical source, the harmonicity, the
9 Different methods for selecting the candidates are analyzed in (Yeh, 2008).

65
4. STATE OF THE ART

smoothness of the spectral envelope, and the synchronous amplitude evolution

of the partials. The harmonicity is estimated by measuring the degree of partial
deviations weighted by their amplitudes. The spectral smoothness is computed
using the mean bandwidth of the FT of the HPS. The spectral centroid is also
considered, favoring those HPS with high energy in the lower partials. Finally,
the synchronicity is measured as the standard deviation of the mean time10 of
the HPS partials. The score function is a weighted sum of these four criteria,
and the weights are optimized using the evolutionary algorithm from Schwefel
(1995) with a large dataset.
The estimation of the number of concurrent sounds is finally done by
iterative score improvement (Chang et al., 2008; Yeh et al., 2006), based on
the explained energy and the improvement of the spectral smoothness. Finally,
a postprocessing stage can be added by tracking the f0 candidates trajectories,
using a high-order HMM and a forward-backward tracking scheme proposed by
Chang et al. (2008).
The joint estimation approach proposed by Emiya et al. (2008b) detects the
onsets in the signal and, for each segment between consecutive onsets, a set of
candidates are selected. Then, the most likely combination of pitches within
each segment is selected using HMM for tracking candidates and a spectral
maximum likelihood method for joint pitch estimation.
The method from Cañadas-Quesada et al. (2009) selects the possible partials
from the magnitude spectral peaks using an adaptive logarithmic threshold. The
f0 candidates are then chosen from the selected partials within a given pitch
range. For each candidate, a harmonic pattern is built in the log-frequency
domain, considering one semitone bandwidth for partial search. Then, all
the possible candidate combinations are generated and explained as a sum of
Gaussian mixture models (GMM). The GMM weights are obtained using the
non-colliding partial magnitudes and the estimated magnitudes11 of colliding
partials. Finally, the combination which explains most of the harmonic content
and maximizes the temporal similarity with the previous winner combinations
is chosen at each frame.

4.2.4 Supervised learning methods

Supervised learning approaches attempt to assign a class to a test sample.
Within this context, the classes are the musical pitches, therefore these methods
are constrained to tuned harmonic sounds, as they estimate the pitch instead
of the fundamental frequency itself.

10 The mean time is an indication of the center of gravity of signal energy (Cohen, 1995). It

can be defined in the frequency domain as the weighted sum of group delays.
11 The colliding partials are estimated by linear interpolation of non-colliding neighbors.

66
 4.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

Figure 4.7: SONIC scheme from Marolt (2004a)

The partial tracking method proposed by Marolt (2004a,b) uses a combi-

nation of the auditory Patterson-Holdsworth gammatone filterbank with the
Meddis hair cell model as a preprocessing stage. Instead of using a correlogram,
a modified version of the Large and Kolen (1994) adaptive oscillators is utilized
to detect periodicities in output channels of the auditory model. There are 88
oscillators with initial frequencies corresponding to the tuned musical pitches.
If the oscillators synchronize with their stimuli (outputs of the auditory model),
then the stimuli are periodic, meaning that partials are present in the input
signal. This scheme can be used to track partials, even in the presence of vibrato
or beating. The model was extended for tracking groups of harmonically related
partials, by using the output of the adaptive oscillators as inputs of neural
networks. A set of 88 neural networks corresponding to the musical pitches
were used, each containing up to 10 oscillators associated to partial frequencies.

The harmonic tracking method from Marolt (2004a,b) was integrated into a
system called SONIC (see Fig. 4.7) to transcribe piano music. The combination
of the auditory model outputs and the partial tracking neural network outputs
are fed into a set of time delay neural networks (TDDN), each one corresponding
to a musical pitch. The system also includes an onset detection stage, which is
implemented with a fully-connected neural network, and a module to detect
repeated notes activations (consecutive notes with the same pitch). The
information about the pitch estimate is complemented with the output of the
repeated note module, yielding the pitch, length and loudness of each note.
The system is constrained to piano transcription, as training samples are piano
sounds.

67
4. STATE OF THE ART

Figure 4.8: HMM smoothed estimation from Poliner and Ellis (2007a) for an
excerpt of Für Elise (Beethoven). The posteriorgram (pitch probabilities as a
function of time) and the HMM smoothed estimation plotted over the ground-
truth labels (light gray) are shown.

Reis et al. (2008c) use genetic algorithms12 for polyphonic piano transcrip-
tion. Basically, a genetic algorithm consist on a set of candidate solutions
(individuals, or chromosomes) which evolve through inheritance, selection,
mutation and crossover until a termination criterion. At each generation, the
quality (fitness) of each chromosome is evaluated, and the best individuals are
chosen to keep evolving. Finally, the best chromosome is selected as the solution.
In the method proposed by Reis et al. (2008c), each chromosome corresponds
to a sequence of note events, where each note has pitch, onset, duration and
intensity. The initialization of the population is based on the observed STFT
peaks. The fitness function for an individual is obtained from the comparison of
the original STFT with the STFT of synthesized versions of the chromosomes
given an instrument. The method is constrained to the a priori knowledge of the
instrument to be synthesized. The system was extended in Reis et al. (2008b)
by combining the genetic algorithm with a memetric algorithm (gene fragment
competition), to improve the quality of the solutions during the evolutionary
process.

12 Genetic algorithms are evolutionary methods based on Darwin natural selection proposed

by Holland (1992).

68
 4.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

Poliner and Ellis (2007a,b) combine Support Vector Machines (SVMs13 ) in

a frame-by-frame spectral analysis with HMMs to refine the estimate. The
proposed approach trains a SVM with spectrograms of synthesized MIDI files.
87 binary note classifiers detect the presence of the notes at a given frame. Each
frame is represented by a 255-element feature vector consisting on normalized
spectral bins. The SVM classification output (called ‘posteriorgram’) is parsed
through a two state (on/off) HMM to increase the temporal coherence in
the pitch estimate. The HMM achieves temporal smoothing (see Fig. 4.8)
by estimating the probability of seeing a particular classifier label given a
true pitch state with the likelihood of each note being ‘on’ according to the
classifiers output. The system is intended for polyphonic piano transcription,
experimentally outperforming the Marolt (2004a,b) results.
SVMs have been also used by Zhou (2006) (Method II), using 88 binary
classifiers, each corresponding to one pitch. The classifier inputs are the peaks
extracted from the RTFI energy spectrum.

4.2.5 Unsupervised learning methods

The goal of non-negative matrix factorization (NMF), first proposed by Lee and
Seung (1999), is to approximate a non-negative matrix Y as a product of two
non-negative matrices W and H, in such a way that the reconstruction error is
minimized:
 
X ≈ WH 4.7 
This method has been used for music transcription, where typically X is the
spectral data, H corresponds to the spectral models (basis functions), and W
are the weightings, i.e., the intensity evolution along time (see Fig. 4.9). This
methodology is suitable for instruments with a fixed spectral profile14 , such as
piano sounds.
There are different ways to design the cost function in order to minimize
the residual. For instance, Cont (2006) assumes that the correct solution for
a given spectrum uses a minimum of templates, i.e., that the solution has the
minimum number of non-zero elements in H. NMF methods have also been used
for music transcription by Plumbley et al. (2002), Smaragdis and Brown (2003),
13 SVMs are supervised learning methods for classification (see (Burges, 1998)). Viewing

input data as sets of vectors in an n-dimensional space, a SVM constructs separating

hyperplanes in that space in such a way that the margins between the data sets are maximized.
14 In the scope of this work, an instrument with a fixed spectral profile is referred when two

notes of that instrument playing the same pitch produce a very similar sound, as it happens
with piano sounds. As an opposite example, a sax can’t be considered to have a fixed spectral
profile, as real sax sounds usually contain varying dynamics and expressive alterations, like
breathing noise, that do not sound in the same way than other notes with the same pitch.

69
4. STATE OF THE ART

Figure 4.9: NMF example from Smaragdis and Brown (2003). The original
score and the obtained values for H and W using 4 components are shown.

Raczynski et al. (2007), Virtanen (2007), Vincent et al. (2007) and Bertin et al.
(2007).
The independent component analysis (ICA), introduced by Comon (1994),
is closely related to the NMF. ICA can express a signal model as x = Wh,
being x and h n-dimensional real vectors, and W a non-singular mixing matrix.
Citing Virtanen (2006), ICA attempts to separate sources by identifying latent
signals that are maximally independent.
As pointed out by Schmidt (2008), the differences between ICA and NMF
are the different constraints placed on the factorizing matrices. In ICA, rows
of W are maximally statistically independent, whereas in NMF all elements
of W and H are non-negative. Both ICA and NMF have been investigated by
Plumbley et al. (2002) and Abdallah and Plumbley (2003a, 2004) for polyphonic
transcription. In the evaluation done by Virtanen (2007) for spectrogram
factorization, the NMF algorithms yielded better separation results than ICA.
These methods have been successfully used for drum transcription (see
(FitzGerald, 2004) and (Virtanen, 2006)), as most percussive sounds have a
fixed spectral profile and they can be modeled using a single component.

4.2.6 Matching Pursuit methods

The Matching Pursuit (MP) algorithm15 from Mallat and Zhang (1993)
approximates a solution for decomposing a signal into linear functions (or atoms)
15 The matching pursuit toolkit (MPTK) from Krstulovic and Gribonval (2006), available

at http://mptk.irisa.fr, provides an efficient implementation of the MP algorithm.

70
 4.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

Figure 4.10: Modified MP algorithm from Leveau et al. (2008) for the extraction
of harmonic atoms.

that are selected from a dictionary. At the first iteration of the algorithm,
the atom which gives the largest inner product with the analyzed signal is
chosen. Then, the contribution of this function is subtracted from the signal
and the process is repeated on the residue. MP minimizes the residual energy
by choosing at each iteration the most correlated atom with the residual. As a
result, the signal is represented as a weighted sum of atoms from the dictionary
plus a residual.
The method proposed by Cañadas-Quesada et al. (2008) is based on
harmonic matching pursuit (HMP) from Gribonval and Bacry (2003). The
HMP is an extension of MP with a dictionary composed by harmonic atoms.
Within this context, a Gabor atom16 can be identified with a partial, and a
harmonic atom is a linear combination of Gabor atoms (i.e., a spectral pattern).
The algorithm from Cañadas-Quesada et al. (2008) extends HMP to avoid
inaccurate decomposition when there are overlapped partials, by maximizing the
smoothness of the spectral envelope for each harmonic atom. The smoothness
maximization algorithm is similar to the one proposed by Klapuri (2003a).
The performance of this method when dealing with harmonically related
simultaneous notes is further described by Ruiz-Reyes et al. (2009).
Leveau et al. (2008) propose a modified MP algorithm which can be applied
to the whole signal, instead of the frame by frame basis. The harmonic atoms
extraction method is shown in Fig. 4.10. Molecules are considered as a group
of several atoms of the same instrument in successive time windows.
16 Gabor atoms are time-frequency atomic signal decompositions proposed by Gabor (1946,

1947). They are obtained by dilating, translating and modulating a mother generating
function.

71
4. STATE OF THE ART

4.2.7 Bayesian models

Citing Davy (2006b), tonal music can be exploited to build a Bayesian model,
that is, a mathematical model embedded into a probabilistic framework that
leads to the simplest model that explains a given waveform. Such models are
also known as generative models because they can be used to generate data by
changing parameters and the noise. Some multiple f0 estimation systems rely on
generative models of the acoustic waveform. Most of these models assume that
the fundamental frequency belongs to a fixed grid, associated to the pitches.
The method proposed by Cemgil et al. (2003, 2006) is based on a generative
model formulated as a dynamical bayesian network. The probabilistic model
assumes harmonic frequency relationships of the partials and exponentially
decaying spectrum envelope from one partial to another. This approach allows
to write many classical noisy sum-of-sines models into a sequential form. The
model relies on sinusoids with damped amplitude and constant frequency.
A piano-roll is inferred from the observation, assigning to each of the grid
frequencies the state ‘mute’ or ‘sound’ at each instant. The algorithm for
estimating the most likely piano-roll is based on EM and Kalman filtering on a
sliding window over the audio signal. This can be considered as a time-domain
method (DFT is not explicitly calculated), which can be used to analyze music
to sample precision, but with a very high computational cost.
Vincent and Rodet (2004) propose a generative model combining a nonlinear
Independent Subspace Analysis (ISA17 ) and factorial HMM. The method is
based on creating specific instrument models based on learning. The spectra of
the instrument sounds are modeled by using the means and variances of partial
amplitudes, the partial frequencies and the residuals. To transcribe a signal, the
spectrum is considered as a sum of spectral models which weights are optimized
using the second order Newton method. The HMM is used for adding temporal
continuity and modeling note duration priors.
Other Bayesian approaches for music transcription are those proposed by
Kashino and Tanaka (1993), Sterian (1999), Walmsley et al. (1999), Raphael
(2002), Kashino and Godsill (2004), Dubois and Davy (2005, 2007), Vincent
and Plumbley (2005), and Davy et al. (2006). For a review on this topic, see
(Cemgil, 2004) and (Davy, 2006b).

4.2.8 Statistical spectral models

Goto (2000) describes a method called PreFEst (see Fig. 4.11) to detect melody
and bass lines in musical signals. The system assumes that the melody and bass
17 ISA combines the multidimensional ICA with invariant feature extraction. Linear ISA

describes the short-time power spectrum of a musical excerpt as a sum of power spectra with
time-varying weights, using a Gaussian noise for modeling the error.

72
 4.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

Figure 4.11: Overview of the system proposed by Goto (2000).

are the most predominant harmonic structures in high and low frequency regions
respectively. First, the STFT is apportioned through a multirate filterbank, and
a set of candidate frequency components are extracted. Then, two bandpass
filters are used to separate the spectral components of the bass and melody.
For each set of filtered frequency components, the method forms a probability
density function (PDF) of the f0 . The observed PDF is considered as being
generated by a weighted mixture of harmonic-structure tone models. The model
parameters are estimated using the EM algorithm. To consider a continuity
of the f0 estimate, the most dominant and stable f0 trajectory is selected,
by tracking peak trajectories in the temporal transition of the fundamental
frequencies PDFs. To do this, a salience detector selects salient promising peaks
in the PDFs, and agents driven by those peaks track their trajectories. The
system works in real time.
Kameoka et al. (2007) propose a method called harmonic temporal struc-
tured clustering (HTC). This approach decomposes the power spectrum time
series into sequential spectral streams (clusters) corresponding to single sources.
This way, the pitch, intensity, onset, duration, and timbre features of each
source are jointly estimated. The input of the system is the observed
signal, characterized by its power spectrogram with log-frequency. The source
model (see Fig. 4.12) assumes smooth power envelopes with decaying partial
amplitudes. Using this model, a goodness of the partitioned cluster is calculated
using the Kullback-Liebler (KL) divergence. The model parameters are
estimated using the expectation-constrained maximization (ECM) algorithm

73
4. STATE OF THE ART

Figure 4.12: HTC spectral model of a single source from Kameoka et al. (2007).

from Meng and Rubin (1993), which is computationally simpler than the EM
algorithm. In the evaluation done by Kameoka et al. (2007), the HTC system
outperformed the PreFEst results.
The method from Li and Wang (2007) is similar to (Ryynänen and Klapuri,
2005) in the sense that the preliminary pitch estimate and the musical pitch
probability transition are integrated into a HMM. However, for pitch estimation,
Li and Wang (2007) use statistical tone models that characterize the spectral
shapes of the instruments. Kernel density estimation is used to build the
instrument models. The method is intended for single instrument transcription.

4.2.9 Blackboard systems

The blackboard systems, introduced by Engelmore and Morgan (1988), have

been also applied to music transcription. The name is based on the idea of
a group of experts standing around a blackboard working together to solve a
problem. The blackboard is a hierarchical data-space, where the hypotheses
are proposed. The experts or knowledge sources typically consist on a set of
rules (a sort of if/then conditions). They develop the hypotheses and remove
the unsupported ones in the blackboard. The scheduler is the third main
component, and it determines the order in which knowledge sources are allowed
to act. The system converges when the knowledge sources are satisfied with the
hypotheses in the blackboard given an error margin.
The general blackboard architecture used in music transcription is similar to
the one proposed by Martin (1996) (see Fig. 4.13). The blackboard hierarchy
is ordered by increasing abstraction, being the input data at the lowest level.
The method from Bello (2000) extends the Martin (1996) architecture by
using top-down processing and a neural network to detect the presence or
absence of a chord. Other blackboard systems for music transcription have
been proposed by Ellis (1996), Monti and Sandler (2002), and Plumbley et al.
(2002), and a review on these methods was published by McKay (2003).

74
 4.2. MULTIPLE FUNDAMENTAL FREQUENCY ESTIMATION

Figure 4.13: Blackboard architecture from Martin (1996).

4.2.10 Database matching

Bello et al. (2002) describes a time-domain approach for piano transcription

using a waveforms database. The goal of this method is to use linear algebra
to decompose the original signal into a sum of signals that are present in the
database. The interesting point is that the waveforms in the database can be
known a priori, but they can also be inferred from the analyzed original signal.

To build the adaptive database, the fundamental frequencies are estimated

using a simplified version of the blackboard method described in (Bello, 2000).
Once the pitches are estimated, the spectral information of each one is used to
resynthesize (using the inverse Fourier transform) an audio signal corresponding
to that pitch. To complete the database, covering all the note range, the signals
of the missing pitches are calculated from the obtained signals using pitch-
shifting by standard phase-vocoder techniques. Once the database is filled, the
time-domain method can estimate the pitches.

The method proposed by Groble (2008) extracts feature vectors from the
spectrum data. These vectors are scored against models computed by scaled
averages of audio samples of piano and guitar from a training set. The training
data consists on sample frames at different note attack levels and distances from
the note onset. The predicted pitches are determined using simple distance
metrics from the observed feature vector to the dataset feature vectors.

75
4. STATE OF THE ART

4.3 Discussion of multiple f0 estimation methods

Multiple f0 estimation is a complex task, and it is very difficult to obtain a single
methodology suitable for all the variety of musical sounds. The advantages and
drawbacks among the different categories previously described are discussed in
this section.
The input data representation is an important choice. However, time-
frequency representations like constant-Q transform, wavelets or filter-banks
(see Sec. 2.1.2) do not seem to provide significant advantages over the STFT
for this task. As it can be seen in Fig. 2.5, and discussed by Hainsworth
(2003) and Yeh (2008), p. 27, multi-resolution schemes do not really solve
the time-frequency main issue. Wavelets sacrifice frequency resolution in the
high frequencies, which can be a drawback for partial search, and constant-Q
and DWT loose temporal precision in the lower frequencies. For these reasons,
most of the multiple f0 estimation methods rely on the STFT.
Most salience methods are computationally efficient, and they provide a
mid-level representation which is useful for f0 estimation. Some of them, are
just enhanced representations which need a posterior methodology to estimate
the fundamental frequencies. Their main drawback is that in some situations
they can lose relevant information or produce spurious components in the signal
transformation process. For instance, SACF performs half wave rectification.
The FT of a half wave rectified signal (see (Klapuri, 2004), Eq. 4.23) consists
of a DC-component, of the original power spectrum scaled down, and of a
convolution of the original power spectrum by itself. The convolution of a
spectrum by itself produces spectral components at the locations that are
multiple of the original spectral intervals, emphasizing partial beating, which
can be very useful for single f0 estimation. However, in polyphonic signals,
all the intervals between partials of different sources (and also between different
fundamental frequencies) generate prominent components at beating frequencies
in the half-wave rectified spectrum, adding spurious data for the posterior
analysis.
Iterative cancellation methods are shown to be efficient, but a wrong f0
estimate may lead to successive errors in an iterative manner and, as pointed
out by Yeh (2008), p. 28, these methods can not estimate well partial overlaps
or beating, as they do not properly consider source interactions. Matching
pursuit methods have similar problems than iterative cancellation methods, as
the contribution of each harmonic atom is subtracted at each iteration before
processing the residual again.
Joint estimation approaches can handle source interactions better than
iterative cancellation methods, but they have high computational costs due to
the evaluation of many possible f0 combinations.

76
 4.4. ONSET DETECTION

Few supervised learning methods have been used as the core methodology for
polyphonic music transcription. Probably, this is because they rely on the data
given in the learning stage, and in polyphonic real music the space of observable
data is huge. However, supervised learning methods have been successfully
applied considering specific instrument transcription (usually, piano) to reduce
the search space. The same issue occurs in database matching methods, as they
also depend on the ground-truth data.
In contrast to supervised approaches, unsupervised learning methods do
not need a priori information about the sources. However, they are suitable
for fixed time-frequency profiles (like piano or drums), but modeling harmonic
sounds with varying harmonic components remains as a challenge (Abdallah
and Plumbley, 2004).
In general, music transcription methods based on Bayesian models are
mathematically complex and they tend to have high computational costs, but
they provide an elegant way of modeling the acoustic signal. Statistical spectral
model methods are also complex but they are computationally efficient.
Blackboard systems are general architectures and they need to rely on other
techniques, like a set of rules (Martin, 1996) or supervised learning (Bello, 2000),
to estimate the fundamental frequencies. However, the blackboard integration
concept provides a promising framework for multiple f0 estimation.

4.4 Onset detection

In this work, onset detection algorithms have been categorized into two main
groups: signal processing methods and machine learning approaches. An
extensive review and comparison of different onset detection systems can be
found in (Bello et al., 2005), (Collins, 2005b), and (Dixon, 2006).

4.4.1 Signal processing methods

Most onset detection methods are based on signal processing techniques, and
they follow the general scheme represented in Fig. 4.14. First, a preprocessing
stage is done to transform the signal into a more convenient representation,
usually in the frequency domain. Then, an onset detection function (ODF),
related to the onset strength at each time frame is defined. A peak picking
algorithm is applied to the detection function, and those peaks over a threshold
are finally identified as onsets.
In the preprocessing stage, most systems convert the signal to the frequency
or complex domain. Besides STFT, a variety of alternative preprocessing
methods for onset detection, like filter-banks (Klapuri, 1999), constant-Q

77
4. STATE OF THE ART

Audio signal

Preprocessing
stage

o(t) 1

0.8

Onset detection 0.6

function 0.4

0.2 θ

0
0 2,32 4,64 6,96 9,28 11,60 13,93 t

Peak picking
and thresholding * * * * * * * * * * * * * *** ** * * *
Figure 4.14: General architecture of most onset detection systems.

transform (Duxbury et al., 2002) or wavelet decomposition (Daudet, 2001) can

be used.
Some examples of onset detection functions are spectral power, spectral
flux, high-frequency content (HFC), phase deviation, weighted phase deviation,
complex deviation, and the Kullback-Liebler divergence. Detailed comparatives
of different ODFs have been done by Dixon (2006), Bello et al. (2005), and
Stowell and Plumbley (2007).
In order to identify the onsets from the ODF, some works just select the
peaks over a constant threshold, whereas others use alternative pick picking
algorithms like a dynamic median threshold (Brossier et al., 2004, Lee and Kuo,
2006), or a moving average (Lacoste and Eck, 2005).
A categorization of the different onset detection algorithms based on signal
processing techniques can be done according to the input data that they use to
obtain the ODF.

Energy-based detection

Many onset detection algorithms attempt to detect abrupt energy variations in

a time-frequency representation.
The method proposed by Klapuri (1999) is based on a psychoacoustic model.
First, in the preprocessing stage, the overall loudness of the signal is normalized,
and a bank composed by 21 filters is used to divide the signal into nearly
critical-bands. The output of each filter is full-wave rectified and decimated, and
amplitude envelopes are extracted for each band. A first order relative difference
function is performed for each amplitude envelope, marking the peaks over a

78
 4.4. ONSET DETECTION

constant threshold as onset components and dropping those components closer

than 50 ms. Then, the onset components from separate bands are combined
using a loudness model, which assigns a value to each onset candidate, yielding
a vector of candidate loudnesses over time. Finally, the peaks of the loudnesses
vector that are over a global threshold are accepted as onsets.
Some approaches are based on the scheme proposed by Klapuri (1999), like
the algorithm implemented by Collins (2005b) for unpitched sounds. In the
preprocessing stage, the STFT is computed, and the spectral bins are combined
according to an ERB filter bank. A decibel transformation is done with equal
loudness contour correction. Peak picking is performed by scoring local peaks
over a seven frame window. A constant threshold is applied to get the onsets.
To improve the temporal resolution, the maximum intensities in small blocks
are taken in the time domain. An onset time is corrected within the blocks
prior to the discovered onset, and this sample position is further corrected to a
nearby zero crossing for smooth segmentation.
Stowell and Plumbley (2007) use adaptive whitening as a preprocessing
stage, normalizing the magnitude of each spectral bin according to a recent
maximum value for that bin, thus obtaining a similar dynamic range over time.
Then, different standard ODFs are evaluated and compared using the whitened
spectral data. The best results were obtained using the complex deviation
function (Dixon, 2006).

Energy and phase-based detection

Only a few systems like (Bello and Sandler, 2003) use the phase information
alone for onset detection. Most methods that take into account the phase, like
(Bello et al., 2004), also consider the energy variations. In the latter work,
abrupt changes of phase and magnitude are combined to get an ODF, and an
adaptive median filter is used for peak picking.
Brossier et al. (2004) use a silence gate to discard onsets in quiet regions,
and implement a set of onset detection functions based on changes in successive
spectral frames, like the high frequency content (HFC) and the complex domain
detection function (CDDF) from Bello et al. (2004). Finally, a median dynamic
threshold is used, considering as onsets those peaks over the threshold. The
system is publicly available as a component of the aubio18 library.
The method proposed by Lee et al. (2007) decomposes the signal into 6 non-
overlapping bands, and applies a forward linear prediction error filter which
coefficients are updated with the least mean squares (LMS) algorithm. After
getting the linear prediction errors for each band, their envelopes are estimated
through rectification, smoothing, and decimation, and they are added together
to get a detection function. Besides this detection function, the phase of the
18 Aubio can be downloaded from http://aubio.org/

79
4. STATE OF THE ART

STFT is also considered, obtaining a phase detection function that is combined

with the linear prediction estimates to yield the onsets.

Transient-based detection

Although not every transient in a signal correspond to an onset, almost all the
musical sounds begin with a transient stage characterized by a non-stationary
part and an abrupt amplitude change.
The method proposed by Röbel (2005) performs the STFT to classify the
spectral peaks into transient peaks, which are potentially part of an attack, and
non transient peaks. This classification is based on the centroid of the time
domain energy of the signal segment related to the analyzed peak. A transient
statistical model determines whether the spectral peaks identified as transients
are produced by background noise or by an onset. The exact onset positions
are determined by estimating the starting time of the transient. The value of
the detection function is normalized, dividing the transient energy by the total
signal energy in the target frame, and a constant threshold is finally applied.
The hybrid approach from Duxbury et al. (2002) aims to detect hard onsets,
considering transient energy variations in the upper frequencies, and soft onsets,
with a FFT-based distance measure at the low frequencies. To do it, the signal
is first split into 5 bands using a constant-Q transform. A transient energy
measure is then used to find transient changes in the upper bands, whereas
the lowest band is analyzed to yield the standard Euclidean distance between
two consecutive FFT vectors. The detection function is based on the difference
between the signal for each band and a smoothed version of itself. The onsets
are detected using an automatic threshold based on a mixture of Gaussians or,
alternatively, on the derivative of the onsets histogram. Finally, onsets across
bands are combined to yield the final estimate through a weighted scheme.

Energy and pitch-based detection

Some methods combine energy information with pitch-related features.

Besides using the RTFI method (Zhou and Mattavelli, 2007) for multiple
f0 estimation, this time-frequency processing technique was also applied by
Zhou et al. (2008) for onset detection. In this method, the RTFI spectrum is
transformed through signal processing techniques into a difference pitch energy
spectrum and a normal pitch energy spectrum. The onset is classified according
to the difference pitch energy spectrum into hard or soft onset. If the onset is
marked as hard, then an energy-based algorithm is selected. This algorithm uses
the difference pitch energy spectrum across all frequency channels to generate
the detection function, which is smoothed by a moving average filter to find the
onsets with a simple peak picking operation. If the onset type is soft, then a

80
 4.4. ONSET DETECTION

pitch-based detection algorithm is selected instead. The pitch-based algorithm

uses the normal pitch energy spectrum to divide the signal into transient and
stable parts, and the transient part is used to locate the onsets by inspecting
the energy changes.

4.4.2 Machine learning methods

Some methods do not follow the general scheme of Fig. 4.14. Instead, they use
machine learning techniques to classify each frame into onset or non-onset.

Supervised learning

The system proposed by Lacoste and Eck (2007) use either STFT and
constant Q transforms in the preprocessing stage. The linear and logarithmic
frequency bins are combined with the phase plane to get the input features for
one or several feed-forward neural networks, which classify frames into onset
or non onset. In the multiple network architecture, the tempo trace is also
estimated and used to condition the probability for each onset. This tempo
trace is computed using the cross-correlation of the onset trace with the onset
trace autocorrelation within a temporal window. A confidence measure that
weights the relative influence of the tempo trace is provided to the network.
Marolt et al. (2002) use a bank of 22 auditory filters to feed a fully connected
network of integrate-and-fire neurons. This network outputs a series of impulses
produced by energy oscillations, indicating the presence of onsets in the input
signal. Due to noise and beating, not all the impulses correspond to onsets.
To decide which impulses are real onsets, a multilayer perceptron trained with
synthesized and real piano recordings is used to yield the final estimates.
Support vector machines have also been used for onset detection by Kapanci
and Pfeffer (2004) and Davy and Godsill (2002) to detect abrupt spectral
changes.

Unsupervised learning

Unsupervised learning techniques like NMF and ICA have been also applied for
onset detection.
Wang et al. (2008) generate the non-negative matrices with the magnitude
spectra of the input data. The basis matrices are the temporal and frequency
patterns. The temporal patterns are used to obtain three alternative detection
functions: a first-order difference function, a psychoacoustically motivated
relative difference function, and a constant-balanced relative difference function.
These ODFs are similarly computed by inspecting the differences of the temporal
patterns.

81
4. STATE OF THE ART

Abdallah and Plumbley (2003b) consider onsets as surprising moments in

the waveform. The detection is based on a probability model of the input, which
generates a moment-by-moment trace of the probability of each observation. In
this method, ICA is used as a conditional probability model, and the probability
assigned by the model to each observation is also used as a form of data for
further analysis.

4.5 Discussion of onset detection methods

Like in multiple f0 estimation, finding a general purpose onset detection
algorithm is an challenging task due to the different onset characteristics.
For instance, smooth onsets need longer temporal information than hard
(percussive) onsets. Citing Tan et al. (2009), due to the inherent variable nature
of onsets resulting from the different types of instruments, no simple algorithm
will be optimal for general onset detection.
Transient and energy are features related to hard onsets, whereas phase and
pitch can be used to detect soft onsets produced by some harmonic sounds.
In onset detection, a good trade-off is hard to achieve, and the combined
approaches considering both hard and soft onsets, i.e. using energy and pitch
information, or energy and phase, are probably the most versatile methods for
this task.

82
 Onset detection using a harmonic
5
filter bank

A novel approach for onset detection is presented in this chapter. The audio
signal is analyzed through a 1/12 octave (one semitone) band-pass filter bank
simulated in the frequency domain, and the temporal derivatives of the filtered
values are used to detect spectral variations related to note onsets.
As previously described, many onset detection methods apply a preprocess-
ing stage by decomposing the signal into multiple frequency bands. In the
perceptually motivated onset detector proposed by Klapuri (1999), a set of
critical band filters is used. Scheirer (1998) uses a six band filter bank, each
one covering roughly one octave range, and Duxbury et al. (2002) performs a
sub-band decomposition of the signal.
The motivation of the proposed approach is based on the characteristics
of most harmonic pitched sounds. The first 5 harmonics of a tuned sound
coincide1 with the frequencies of other pitches in the equal temperament (see
Fig. 2.16). Other characteristic of these sounds is that usually most of their
energy is concentrated in the first harmonics. A one semitone filter bank is
composed by a set of triangular filters, which center frequencies coincide with
the musical pitches (see Fig. 5.1).
In the sustain and release stages of a sound, there can be slight variations
in the intensity (tremolo) and the frequency (vibrato) of the harmonics. For
instance, a harmonic peak at the frequency bin k in a given frame can be
shifted to the position k + 1 in the following frame. In this scenario, direct
spectra comparison, like spectral flux (see Eq. 2.14), may yield false positives,
as intensity differences are detected. Using this musically motivated filter bank,
the value of the band which center is close to k will be similar in both frames,
avoiding a false detection.

1 with slight deviations when inharmonicity is present.

83
5. ONSET DETECTION

... ...

C3 C!3 D3 D!3 E3 F3 F !3 G3 frequency

...
b1 ... bi bB Energy in
each band

Figure 5.1: One semitone filter bank.

Therefore, by using one semitone filters, the effects of subtle spectrum

variations produced during the sustain and release stages of a note are
minimized, whereas in the attack the filtered amplitudes increase significantly,
as most of the energy of the partials is be concentrated in the center frequencies
of the semitone bands. This way, the system is specially sensitive to frequency
variations that are larger than one semitone. Therefore, the method is primarily
based on energy changes, but also considering the harmonic properties of the
sounds.
The proposed approach has been evaluated and compared to other works.
Some contents of this chapter have been published in (Pertusa et al., 2005)
and (Pertusa and Iñesta, 2009). The algorithm, developed in C++, has been
publicly released2 for research purposes.

5.1 Methodology
For detecting the beginnings of the notes in a musical signal, the method
analyzes the spectrum information across one semitone filter bank, computing
the band differences in time to obtain a detection function. Peaks in this
function are extracted, and those which values are over a threshold are
considered as onsets.

5.1.1 Preprocessing
From a digital audio signal, the STFT is computed, providing its magnitude
spectrogram. A Hanning window with 92.9 ms length is used, with a 46.4 ms
hop size. With these values, the temporal resolution achieved is ∆t = 46.4 ms,
and the spectral resolution is ∆f = 10.77 Hz.
2 http://grfia.dlsi.ua.es/cm/worklines/pertusa/onset/pertusa_onset.tgz

84
 5.1. METHODOLOGY

Using a 1/12 octave filter bank, the filter corresponding to the pitch G]0
has a center frequency of 51.91 Hz, and the fundamental frequency of the next
pitch, A0 , is 55.00 Hz, therefore this spectral resolution is not enough to build
the lower filters. Zero padding was used to get more points in the spectrum.
Using a zero padding factor z = 4, three additional windows with all samples set
to zero were appended at the end of each frame before doing the STFT. With
this technique, a frequency resolution ∆f = 10.77/4 = 2.69 Hz is eventually
obtained.
At each frame, the spectrum is apportioned among a one semitone filter bank
to produce the corresponding filtered values. The filter bank comprises from 52
Hz (pitch G]0 ) to the Nyquist frequency to cover all the harmonic range. When
fs = 22, 050 Hz, B = 94 filters are used3 , which center frequencies correspond to
the fundamental frequencies of the 94 notes in that range. The filtered output
at each frame is a vector b with B elements (b ∈ B ). R
 
b = {b1 , b2 , . . . , bi , . . . , bB } 5.1 
Each value bi is obtained from the frequency response Hi of the correspond-
ing filter i with the spectrum. The Eq. 5.2 is used4 to compute the filtered
values:
v
 
uK−1
uX
bi = t (|X[k]| · |Hi [k]|)2 5.2 
k=0

5.1.2 Onset detection functions

Two onset detection functions have been used. The first one, called o[t], can
be used for percussive (hard) onsets, whereas an alternative ODF õ[t] has been
proposed for sounds with smooth (soft) onsets.

Onset detection function for hard onsets (o[t])

Like in other onset detection methods, as (Bilmes, 1993), (Goto and Muraoka,
1995, 1996), and (Scheirer, 1998), a first order derivative function is used to
pick potential onset candidates. In the proposed approach, the derivative c[t] is
computed for each filter i.

d  
ci [t] = bi [t] 5.3 
dt

3 When fs = 44, 100 Hz, there are B = 106 bands.

4 This equation was selected instead of Eq. 2.9 since it experimentally yielded better results
for this task.

85
5. ONSET DETECTION

o(t) 1

0.8

0.6

0.4

0.2 θ

0
0 2,32 4,64 6,96 9,28 11,60 13,93 t

Figure 5.2: Example of the onset detection function o[t] for a piano melody,
RWC-MDB-C-2001 No. 27 from Goto (2003), RWC database.

The values for each filter must be combined to yield the onsets. In order to
detect only the beginnings of the events, the positive first order derivatives of
all the bands are summed at each time, whereas negative derivatives, which can
be associated with offsets, are discarded:

B
X  
a[t] = max {0, ci [t]}. 5.4 
i=1

To normalize the onset detection function, the overall energy s[t] is also
computed (note that a[t] < s[t]):

B
X  
s[t] = bi [t] 5.5 
i=1

The sum of the positive derivatives a[t] is divided by the sum of the filtered
values s[t] to compute a relative difference. Therefore, the onset detection
function o[t] ∈ [0, 1] is:

B
X
max {0, ci [t]}
a[t] i=1
 
o[t] = = B 5.6 
s[t] X
bi [t]
i=1

Fig. 5.2 shows an example of the onset detection function o[t] for a piano
excerpt, where all the peaks over the threshold θ were correctly detected onsets.

86
 5.1. METHODOLOGY

Onset detection function for soft onsets (õ[t])

The previous methodology yields good results for instruments that have a sharp
attack, like a piano or a guitar. But for instruments with a very smooth attack,
like violins, more frames should be considered. For these sounds, Eq. 5.3 can be
replaced by:

C
X  
c̃i [t] = j · (bi [t + j] − bi [t − j]) , 5.7 
j=1

being C the number of adjacent frames considered. This expression is based

on the method proposed by Young et al. (2000), Eq. 5.16, to enhance the
performance of a speech recognition system.
Using this scheme, the difference is centered on each particular frame, thus
two side difference (with C = 1) is used instead of the frame itself. When using
C = 2, the derivative is calculated for a longer period. Using Eq. 5.7,

B
X  
ã[t] = max {0, c̃i [t]} 5.8 
i=1

With these equations, Eq. 5.5 must be replaced by Eq. 5.9 to normalize õ[t]
into the range [0, 1]:

B X
X C  
s̃[t] = j · bi [t + j] 5.9 
i=1 j=1

Therefore, õ[t] is calculated as:

B
X
max {0, c̃i [t]}
ã[t] i=1
 
õ[t] = = B C 5.10 
s̃[t] XX
j · bi [t + j]
i=1 j=1

An example of the onset detection function for a violin sound is shown in

Fig. 5.3, without considering additional frames (a), with C = 1 (b), and with
C = 2 (c).
Note that the higher C is, the lower the temporal precision is for detecting
onsets, but the success rate can improve in some instruments which onsets are
difficult to detect. For an adequate detection, the notes must have a duration
l ≥ ∆t(C + 1). With the used parameters, if C = 2 then l = 139.2 ms, so this
methodology is not suitable for very rapid onsets5 .
5 139 ms is the duration of a sixteenth note when tempo = 107 bpm.

87
5. ONSET DETECTION

o[t]
o(t) 1

0.8

0.6

0.4

0.2 θ

0
0 2,32 4,64 6,96 9,28 11,60 13,93 t
õ[t]
o(t) 1

0.8

0.6

0.4

0.2 θ

0
0 2,32 4,64 6,96 9,28 11,60 13,93 t
o(t)
õ[t] 1

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2 θ

0.1

0
0 2,32 4,64 6,96 9,28 11,60 13,93 t

Figure 5.3: Onset detection function for a polyphonic violin song (RWC-MDB-
C-2001 No. 36 from Goto (2003), RWC database). (a) o[t]; (b) õ[t], with C = 1;
(c) õ[t], with C = 2. With C = 2, all the onsets were successfully detected
except by one, which is marked with a circle.

88
 5.2. EVALUATION WITH THE ODB DATABASE

5.1.3 Peak detection and thresholding

The last stage is to extract the onsets from the onset detection function. Peaks at
time t are identified in the onset detection function when o[t−1] < o[t] > o[t+1],
and those peaks over a fixed threshold o[t] > θ are considered as onsets. Two
consecutive peaks can not be detected, therefore the minimum temporal distance
between two onsets is 2∆t = 92.8 ms. A silence threshold µ is also introduced
to avoid false positive onsets in quiet regions, in such a way than if s[t] < µ,
then o[t] = 0. The same peak detection and thresholding procedure is applied
for õ[t].
The silence gate µ is only activated when silences occur, or when the
considered frame contains very low energy, therefore it is not a critical
parameter. The precision/recall deviation can be controlled through the
threshold θ.

5.2 Evaluation with the ODB database

To evaluate the method and set up the thresholds, a test database called ODB
(onset detection database) has been built using a set of real recordings6 . The
ground-truth onset positions were marked and reviewed (see Fig. 5.4) using the
speech filling system (SFS7 ) software. The ODB data set contains a number of
sounds selected from the Goto (2003) RWC database, plus other real recordings.
The songs were selected to cover a relatively wide range of instruments and
musical genres.
The ODB database, with the audio signals and the labeled onsets, has been
publicly released8 for research purposes. An onset detection evaluation software,
which is also available9 has been developed to compare the detected onsets with
the ground-truth. This algorithm computes the number of correct detections,
false positives, and false negatives, considering a 50 ms error margin.

5.2.1 Results using o[t]

As previously described, the system has two free parameters; the silence gate
threshold µ, to avoid false positives when the signal level is very low, and the
onset detection threshold θ, which controls the precision/recall deviation of the
ODF. The method was evaluated with the ODB database to set an appropriate
value for θ. The results are shown in Fig. 5.5. A good compromise between
precision and recall was obtained using θ = 0.18, with µ = 70.
6 Thanks to Jason Box for labeling this data set.
7 http://www.phon.ucl.ac.uk/resource/sfs/
8 http://grfia.dlsi.ua.es/cm/worklines/pertusa/onset/ODB
9 http://grfia.dlsi.ua.es/cm/worklines/pertusa/onset/evaluator

89
5. ONSET DETECTION

Figure 5.4: Onsets from RWC-MDB-C-2001 No. 27 from Goto (2003), RWC
database, labeled with speech filling system (SFS).

0.9

0.8

0.7

0.6

0.5

0.4

Precision
Recall
0.3
0 0.1 0.2 0.3 0.4 0.5
θ

Figure 5.5: Onset detection (o[t]) precision and recall curves in function of the
threshold θ, using a constant value for the silence threshold µ = 70.

90
 5.2. EVALUATION WITH THE ODB DATABASE

Reference Content OK FP FN M D Pr % Re % F-m %

RWC-C02 classic 64 37 43 2 2 66.37 59.81 61.54
RWC-C03 classic 25 41 31 0 0 37.88 44.64 40.98
RWC-C26 piano 36 1 0 0 0 97.30 100.0 98.63
RWC-C27 piano 210 0 13 0 0 100.0 94.17 97.00
RWC-C36 violin 45 19 0 0 0 70.31 100.0 82.57
RWC-C38 violin 165 13 29 0 0 92.70 85.05 88.71
RWC-J01 piano 96 6 17 0 0 94.12 84.96 89.30
RWC-G08 rock 62 17 8 0 0 78.48 88.57 83.22
2-artificial soul 117 35 16 1 1 76.97 87.97 82.11
2-uncle mean jazz 157 10 21 0 0 94.01 88.20 91.01
3-long gone rock 135 53 8 0 0 71.81 94.41 81.57
3-you think too much jazz 160 10 25 0 0 94.12 86.49 90.14
6-three rock 138 27 17 0 0 83.64 89.03 86.25
8-ambrielb electro 111 33 36 0 0 77.08 75.51 76.29
15-tamerlano opera 127 41 12 0 0 75.60 91.37 82.74
25-rujero guitar 92 17 2 0 0 84.40 97.87 90.64
Its alright for you rock 83 1 1 0 0 98.81 98.81 98.81
Tiersen 11 bells 37 36 1 0 0 50.68 97.37 66.67
Realorgan3 organ 13 10 2 0 0 56.52 86.67 68.42
Total 1873 406 282 3 3 82.19 86.91 84.48

Table 5.1: Onset detection results using the proposed database (ODB). The
table shows the number of correctly detected onsets (OK), false positives (FP),
false negatives (FN), merged onsets (M), doubled onsets (D), precision (P),
recall (R), and F-measure (F-m).

The detailed results using o[t] with these thresholds can be seen in Tab. 5.1.
The overall F-measure achieved was 84.48%.
In order to get a perceptual evaluation of the results, once performed the
onset detection, new audio files10 were generated using CSound by adding to the
original waveform a click sound in the positions where the onsets were detected.

Comparison with other approaches

In order to compare the method with other approaches, two publicly available
onset detection algorithms were evaluated using the ODB database. The
experiments were done comparing the onset times obtained by BeatRoot11 and
aubio12 with the ground-truth onsets of the ODB database using the evaluation
methodology previously described.
BeatRoot, introduced by Dixon (2006), is a software package for beat
tracking, tempo estimation and onset detection. To evaluate the method, the
onset times were obtained using the BeatRoot-0.5.6 default parameters with the
following command:

java -jar beatroot-0.5.6.jar -o onsets.txt -O input.wav

10 http://grfia.dlsi.ua.es/cm/worklines/pertusa/onset/generated_sounds_ODB
11 http://www.elec.qmul.ac.uk/people/simond/beatroot/index.html
12 http://aubio.org/

91
5. ONSET DETECTION

System OK FP FN M D Pr % Re % F-m %
Pertusa et al. (2005) 1873 406 282 3 3 82.19 86.91 84.48
Brossier (2005) - aubio 1828 608 327 79 80 75.04 84.83 79.63
Dixon (2006) - BeatRoot 1526 778 629 21 21 66.23 70.81 68.45

Table 5.2: Comparison with other methods using the ODB database and o[t].

In the method from Dixon (2006), different onset detection functions based
on spectral flux, phase deviation, and complex domain13 can be selected.
The onset detection function values are normalized and a simple peak picking
algorithm is used to get the onset times.
Aubio is the implementation of the algorithm proposed by Brossier (2005),
submitted to the MIREX 2005 contest and previously described in Sec. 4.4.
Like in BeatRoot, the default parameters were used for the evaluation of this
method:

aubioonset -i input.wav > onsets.txt

The results (Tab. 5.2) show that the proposed method outperforms these
two approaches using the ODB data set.

5.2.2 Results using õ[t]

For pitched instruments with a non-percussive onset, like a violin or a church
organ, more frames should be considered due to the longer attack stage. In the
ODB test set, these kind of sounds only appear in five songs. The evaluation
results for these melodies using C = 2 and θ = 0.18 are shown in Tab. 5.3. The
results with C = 1 were quite similar to those obtained without considering
additional frames, therefore they are not shown in the table.
In the songs missing in the table, i.e., those with hard onsets, the F-measure
importantly decreases as the number of frames increases, mainly due to the
worse temporal resolution. Using õ[t] for the non-percussive onset audio files,
the system only yielded better results for the RWC-C03 and RWC-C36 melodies,
as shown in Tab. 5.3.
Therefore, this methodology is only suitable for very specific musical pieces,
and it is not adequate for a general purpose onset detection method, yielding
better results only in a very small subset of non-percussive pitched sounds.

13 Based on the estimation of the expected amplitude and phase of the current bin according

to the previous two bins.

92
 5.3. MIREX EVALUATION

Reference OK FP FN M D Pr % Re % F-m %
RWC-C02 47 35 60 0 0 57.32 43.93 49.74
RWC-C03 25 38 31 0 0 39.68 44.64 42.02
RWC-C36 45 5 0 0 0 90.00 100.00 94.74
RWC-C38 101 45 93 5 5 69.18 52.06 59.41
Realorgan3 11 13 4 0 0 45.83 73.33 56.41

Table 5.3: Results using õ[t], with C = 2.

5.3 MIREX evaluation

In order to compare the proposed method with other approaches using an
independent and large data set, it was evaluated in the MIREX (2005) onset
detection contest. As previously discussed, õ[t] is not suitable for general
purpose onset detection and the results were not satisfactory for most sounds,
therefore the algorithm was submitted only with the o[t] function.
For this challenge, the method was implemented using D2K and M2K
modules. The D2K (data to knowledge) toolkit14 is a modular environment
for data mining, whereas M2K15 is a D2K module set for music information
retrieval applications developed by the IMIRSEL16 project.
The D2K itinerary submitted to MIREX (2005) is shown in Fig. 5.6.
Unfortunately, there was a bug in the code which caused a very low F-measure
in the evaluation. The problem was that the mean deviation from the ground-
truth onsets was of −22 ms, because all the onset times were reported as the
beginning time of the frame, instead of the center.
As the method was not properly evaluated in MIREX (2005), the algorithm
was implemented again in C++ and resubmitted to the MIREX (2009) onset
evaluation contest. The MIREX (2009) data set and the evaluation metrics are
exactly the same used in previous contests (MIREX, 2005, 2006, 2007), therefore
the results are comparable with those obtained in previous editions.

5.3.1 Methods submitted to MIREX 2009

Besides the proposed method, three algorithms were evaluated in the MIREX
(2009) onset detection contest. These algorithms are briefly described next.
The onset detection function from Tzanetakis (2009) is based on the half-
wave rectified spectral flux. It uses a pick picking algorithm to find local maxima
in consecutive frames and a threshold relative to the local mean. To reduce the

14 http://alg.ncsa.uiuc.edu/do/tools/d2k
15 Music to knowledge, http://www.music-ir.org/evaluation/m2k/
16 International Music Information Retrieval Systems Evaluation Laboratory.

93
5. ONSET DETECTION

Figure 5.6: D2K itinerary of the system submitted to MIREX (2005).

number of false positives, the ODF is smoothed using a Butterworth filter17

both in the forward and reverse directions to avoid phase distortions.
The Röbel (2009) method is based on the classification of spectral peaks into
transients and non-transients. The algorithm is very similar to that of Röbel
(2005)18 , as it analyzes the transient peak probability in the different spectral
bands. The main difference from the previous approach is the use of bands
with harmonically related center frequencies, instead of continuous frequency
regions, to better detect pitched transients. Five versions were submitted to
the MIREX (2009) onset detection contest. The algorithms labeled as 12 nhd
and 16 nhd were trained with the same data set used in (Röbel, 2005), whereas
the algorithms 7 hd, 10 hd and 19 hdc were trained with additional synthesized
sounds.
The method proposed by Tan et al. (2009) integrates energy-based and pitch-
based detections. First, the system attempts to distinguish between pitched
percussive onsets (like pianos), non-percussive onsets (like brasses or winds), and
unpitched onsets (like drums). This is done by detecting the average bandwidth
of abrupt energy changes (percussiveness) and estimating the pitch tuning in a
similar way than in (Zhu and Kankanhalli, 2006). A musical piece is classified
into one of these three categories using heuristic rules over the percussiveness
and the pitch tuning measures.
17 A Butterworth (1930) filter is designed to have a frequency response which is maximally

flat in the passband.

18 See Pag. 80 for a description of this method.

94
 5.3. MIREX EVALUATION

Participant OK FP FN M D Pr % Re % F-m %
Röbel (2009) 10 hd 7015 1231 2340 161 133 85.00 79.19 79.60
Röbel (2009) 7 hd 7560 2736 1795 188 257 81.32 83.30 79.00
Röbel (2009) 19 hdc 7339 2367 2016 185 212 80.56 81.88 78.31
Pertusa and Iñesta (2009) 6861 2232 2494 196 10 79.99 77.50 76.79
Röbel (2009) 16 nhd 6426 846 2929 148 183 86.39 73.62 76.48
Röbel (2009) 12 nhd 6440 901 2915 145 198 85.96 73.15 76.10
Tan et al. (2009) 1 6882 2224 2473 157 308 75.67 76.97 74.43
Tan et al. (2009) 2 6588 1976 2767 152 266 78.28 74.58 73.38
Tan et al. (2009) 3 5961 1703 3394 146 285 79.61 68.97 68.63
Tan et al. (2009) 5 7816 5502 1539 84 1540 62.88 83.69 68.23
Tan et al. (2009) 4 5953 1843 3402 135 345 78.98 68.91 67.94
Tzanetakis (2009) 5053 2836 4302 162 46 67.01 59.91 59.54

Table 5.4: Overall MIREX 2009 onset detection results ordered by F-measure.
The precision, recall and F-measure are averaged. The highest F-measure was
obtained using θ = 0.25.

If the musical excerpt is classified as unpitched, the onset detection is

based only on energy processing. If it is set as pitched (percussive or non-
percussive), both energy processing and pitch processing are combined. The
energy-based processing computes the spectral differences from two consecutive
frames and applies an adaptive threshold. The pitch-based processing computes
the chromagram and looks for changes in the strongest base pitch class and
dominant harmonics pitch class pair. Adjacent time frames with the same pitch
content are grouped into the same cluster, and the clusters indicate regions
belonging to the same note. In the percussive pitched class, energy changes
have higher weights than pitch changes, whereas in the non-percussive category
the pitch changes are considered to be more relevant.

5.3.2 MIREX 2009 onset detection results

The method was evaluated in the MIREX (2009) onset detection contest using
different values of θ ∈ [0.1, 0.3]. The overall results using the best parameters
for each algorithm are shown in Tab. 5.4. The proposed method yielded a good
average F-measure with a very low computational cost (see Tab. 5.5).
The Tab. 5.6 shows the values of θ that yielded the best results for each
sound category compared with the highest overall average F-measure among
the evaluated methods. The proposed approach achieved the highest average
F-measure for the brass, drums and plucked strings categories, characterized by
hard onsets.
Complex sounds are mixtures of unpitched and pitched sounds, including
singing voice. In this category, the method achieved a lower F-measure than
using poly-pitched and drum sounds probably due to the presence of singing
voice. In general, this algorithm is not suitable for singing voice, which is
usually not perfectly tuned and tends to have partials shifting in frequency
across different semitones, causing many false positives.

95
5. ONSET DETECTION

Participant Runtime (hh:mm)

Pertusa and Iñesta (2009) 00:01
Tzanetakis (2009) 00:01
Röbel (2009) 12 nhd 00:02
Röbel (2009) 16 nhd 00:02
Röbel (2009) 7 hd 00:03
Röbel (2009) 19 hdc 00:03
Röbel (2009) 10 hd 00:04
Tan et al. (2009) 1 01:57
Tan et al. (2009) 2 01:57
Tan et al. (2009) 3 01:57
Tan et al. (2009) 4 01:57
Tan et al. (2009) 5 01:57

Table 5.5: MIREX 2009 onset detection runtimes.

Class Files θ Pr % Re % F-m % Best F-m %

Complex 15 0.19 68.97 70.25 68.51 74.82
Poly-pitched 10 0.20 93.58 88.27 90.36 91.56
Solo bars & bells 4 0.26 79.69 80.12 77.34 99.42
Solo brass 2 0.24 77.69 83.68 79.88 79.88
Solo drum 30 0.21 94.00 88.93 90.68 90.68
Solo plucked strings 9 0.28 91.25 89.56 90.11 90.11
Solo singing voice 5 0.30 15.17 46.12 22.63 51.17
Solo sustained strings 6 0.24 58.06 60.92 55.47 64.01
Solo winds 4 0.28 54.97 70.17 60.15 75.33

Table 5.6: Detailed MIREX 2009 onset detection results for the proposed
method with the best θ for each class. The precision, recall and F-measure
are averaged. The best F-measure among the evaluated methods is also shown.

Participant Params Pr % Re % F-m %

Röbel (2009) 19 hdc 0.34 92.07 92.02 91.56
Röbel (2009) 7 hd 0.34 91.70 92.31 91.54
Pertusa and Iñesta (2009) 0.20 93.58 88.27 90.36
Röbel (2009) 19 hd 0.52 93.02 87.79 89.20
Röbel (2009) 16 nhd 0.43 98.51 80.21 87.26
Röbel (2009) 12 nhd 0.49 96.12 80.97 87.11
Tan et al. (2009) 1 N/A 85.94 83.11 83.12
Tan et al. (2009) 2 N/A 85.94 83.11 83.12
Tan et al. (2009) 3 N/A 89.61 70.73 74.17
Tan et al. (2009) 4 N/A 89.23 70.32 73.77
Tan et al. (2009) 5 N/A 61.33 90.01 68.66
Tzanetakis (2009) N/A 71.91 66.43 67.41

Table 5.7: MIREX 2009 poly-pitched results ordered by average F-measure.

96
 5.3. MIREX EVALUATION

Complex Solo drum Solo winds

Poly pitched Solo plucked strings Total
Solo bars and bells Solo singing voice
Solo brass Solo sustained strings
F-m

Figure 5.7: MIREX 2009 onset detection F-measure respect to the threshold θ
for the different sound classes using the proposed method.

The proposed methodology is primarily intended for detecting pitch changes,

therefore the poly-pitched results (see Tab. 5.7) are of special interest. For
this class of sounds, the F-measure was close to the best. These results are
satisfactory, given that the Röbel (2009) and Tan et al. (2009) approaches are
also oriented to the onset detection of pitched sounds.
As expected, the method performs slightly worse with sounds with non-
percussive attacks, like sustained strings and winds. For instance, for sax
sounds two onsets are usually detected by the system; one when the transient
(breathing) begins, and other when the pitch is reached. Portamentos19 are
also a problem for the proposed method, and they usually occur in these kind
of sounds, and also in singing voice. A strong portamento produces that a new
onset is detected each time that a semitone is reached, like it would happen
with a glissando. This is not a drawback for multiple pitch estimation systems,
but it may yield some false positive onsets. Therefore, for detecting the onsets
of these sounds, it is probably more adequate to identify the transients rather
than the pitch changes.
19 A portamento is a continuous and smooth frequency slide between two pitches. A glissando

is a portamento which moves in discrete steps corresponding to pitches. For instance, a

glissando can be played with a piano, but this instrument is unable to play a portamento.
Violins can produce portamentos, although they can also generate glissandos.

97
5. ONSET DETECTION

Bars and bells have percussive onsets and they are typically pitched,
although most of these sounds are inharmonic. Therefore, their energy may
not be concentrated in the central frequencies of the one semitone bands. In
the proposed method, when this happens and the harmonics slightly oscillate
in frequency, they can easily reach adjacent bands, causing some false positives.
Anyway, it is difficult to derive conclusions for this class of sounds, as only 4
files were used for the evaluation and the MIREX data sets are not publicly
available.
Interestingly, the proposed approach also yielded good results with unpitched
sounds, and it obtained the highest F-measure in solo-drum excerpts among all
the evaluated methods.
The best threshold value for poly-pitched and complex sounds was around
θ = 0.20, which coincides with the threshold experimentally obtained with the
ODB database. Using this threshold, the overall F-measure is only 1% lower
(see Fig. 5.7) than with the best threshold θ = 0.25 for the whole MIREX data
set, therefore the differences are not significant.

5.4 Conclusions
An efficient novel approach for onset detection has been described in this
chapter. In the preprocessing stage, the spectrogram is computed and
apportioned through a one semitone filter bank. The onset detection function
is the normalized sum of temporal derivatives for each band, and those peaks
in the detection function over a constant threshold are identified as onsets.
A simple variation has been proposed, considering adjacent frames in order
to improve the accuracy for non-percussive pitched onsets. In most situations,
õ[t] yields lower results than without considering additional frames, therefore it
is only suitable for a few specific sounds.
The method has been evaluated and compared with other works in the
MIREX (2009) audio onset detection contest. Although the system is mainly
designed for tuned pitched sounds, the results are competitive for most timbral
categories, except for speech or inharmonic pitched sounds.
As the abrupt harmonic variations produced at the beginning of the notes are
emphasized and those produced in the sustain stage are minimized, the system
performs reasonably well against smooth vibratos lower than one semitone.
Therefore, o[t] is suitable for percussive harmonic onsets, but it is also robust
to frequency variations in the sustained sounds.
When a portamento occurs, the system usually detects a new onset when
the f0 increases or decreases more than one semitone, resulting in some false
positives. However, this is not a drawback if the method is used for multiple
pitch estimation.

98
 5.4. CONCLUSIONS

As future work, a first categorization of the sounds could be done, like in

(Tan et al., 2009). If the song belongs to the non-percussive category, õ[t] could
be used instead, dynamically adjusting the value of C according to the degree
of percussiveness.
An adaptive filter bank could also be included for mistuned harmonic sounds,
determining the tuning like in (Zhu and Kankanhalli, 2006) and shifting the
center frequencies of the bands according to the tuning deviation.
Other aspect that might be improved is the temporal resolution, by changing
the window length, the hop size, or implementing a similar scheme than in
(Collins, 2005a) to refine the onset position after the detection.

99
 Multiple pitch estimation using
6
supervised learning methods

A simple first approach for multiple pitch estimation in monotimbral1 mixtures

is presented in this chapter. Some of the following contents have been published
in (Pertusa, 2003; Pertusa and Iñesta, 2004; Pertusa and Iñesta, 2005).
The hypothesis is that a supervised method can learn a pattern from a
given timbre. The system is trained using spectrograms apportioned into a
one semitone filter bank and the information about the ground-truth pitches.
After the learning stage, the supervised algorithm can detect the pitches
as occurrences of that pattern in the spectrogram, even in the presence
of polyphony interferences (harmonic overlap). Time-delay neural networks
(TDNN) and k-nearest neighbours (kNN) have been evaluated and compared
for pitch classification.
The main drawback of supervised learning techniques is that they depend on
the training data. The amount of timbral variety and pitch combinations in the
training stage can condition the success rate. As it is difficult to get real musical
data aligned with the ground-truth pitches, synthesized songs have been used
in a simplified scenario, obtaining promising results for tuned synthetic sounds
with fixed spectral profiles and constant amplitude envelopes. The multiple
f0 estimation of real music without any a-priori information is addressed in
Chapter 7.
As previously discussed in Sec. 4.2.4, different supervised techniques have
been used for multiple pitch estimation, most of them for piano sounds. It should
be mentioned that the proposed methods described in the present chapter were
published in 2004, and all the supervised learning approaches cited in 4.2.4 are
posterior, except from the Marolt (2001) work which was the basis for (Marolt,
2004a,b).

1 Sounds produced by a single timbre.

101
6. MULTIPLE PITCH ESTIMATION USING SUPERVISED LEARNING

Besides the above-mentioned methods for multiple pitch estimation, many

supervised learning techniques have been used for single pitch estimation of
speech and musical monophonic signals. An earlier work which is similar to
the presented approach was proposed by Taylor and Greenhough (1993) for
single pitch classification. The inputs of this method were the spectra of signals
sampled at 8 kHz mapped on a distribution of one semitone bins, like in (Sano
and Jenkins, 1989). Then, an adaptive resonance theory neural network called
ARTMAP (Carpenter et al., 1991) was used for pitch classification, obtaining
high success rates.
Time delay neural networks have also been successfully used in speech
recognition problems (Waibel, 1989, Bourlard and Morgan, 1992), and for
multiple pitch estimation by Marolt (2004a). In the latter work, TDNNs
were compared with multilayer perceptrons, radial basis function networks2 and
Elman partially recurrent networks3 . Experimentally, TDNNs achieved the best
results among the evaluated methods.

6.1 Preprocessing
Supervised learning techniques require a set of input features aligned with
the desired outputs. In the proposed approach, a frame by frame analysis
is performed, building input-output pairs at each frame. The input data are
spectral features, whereas the outputs consist on the ground-truth pitches. The
details of the input and output data and their construction are described in this
section.

6.1.1 Construction of the input-output pairs

In order to increase the computational efficiency and to reduce the amount
of spurious information in the input vectors, it is important to reduce the
dimensionality of the feature set. If all the spectral bins were used, there would
be 2048 or 4096 input features. Some of them are unnecessary, thus they can
complicate the learning process and increase the computational complexity.
In the proposed method, like in the onset detection approach previously
described, a one semitone filter bank is used to compress in some way the
spectral information. As most of the spectral energy in harmonic sounds usually
coincide with the center of the one semitone filters, this kind of filter bank can
be used to represent the energy of the prominent harmonics, keeping the main
structure of a harmonic pattern.
2 A radial basis function network is an artificial neural network that uses radial basis

functions as activation functions.

3 The Elman network is a variation on the multilayer perceptron, with the addition of a set

of context units in the input layer connected with the hidden layer units.

102
 6.2. SUPERVISED METHODS

Input data

The training data set consists of musical audio files at fs = 22, 050 Hz
synthesized from MIDI sequences. The STFT of the each musical piece is
computed, providing the magnitude spectrogram using a 93 ms Hanning window
with a 46.4 ms hop size. With these parameters, the time resolution for
the spectral analysis is ∆t = 46.4 ms, and the highest possible frequency is
fs /2 = 11025 Hz, which is high enough to cover the range of useful pitches.
Like in the onset detection method, zero padding has been used to build the
lower filters.
The same way as described in Chapter 5 for onset detection, the spectral
values at each frame are apportioned into B = 94 bands using a one semitone
filter bank ranging from 50 Hz (G]0 ) to fs /2, almost eight octaves, yielding a
vector of filtered values b[t] at each frame.
These values are converted into decibels and set as attenuations from the
maximum amplitude, which is 96 dB4 with respect to quantization noise. In
order to remove noise and low intensity components at each frame, a threshold
ξ is applied for each band in such a way that, if bi [t] < ξ, then bi [t] = ξ. This
threshold was empirically established at ξ = −45 dB. This way, the input data
is within the range b[t] ∈ [ξ, 0]B .
Information about adjacent frames is also considered to feed the classifiers.
For each frame at time t, the input is a set of spectral features {b[t + j]} for
j ∈ [−m, +n], being m and n the number of spectral frames considered before
and after the frame t, respectively.

Output data

For each MIDI file, a binary digital piano-roll (BDP) is obtained to get the
active pitches (desired output) at each frame. A BDP is a matrix where each
row corresponds to a frame and each column corresponds to a MIDI pitch (see
Fig. 6.1). Therefore, at each frame t, n + m + 1 input vectors b[t + j] for
j ∈ [−m, +n] and a vector of pitches ν[t] ∈ {0, 1}B are shown to the supervised
method during the training stage.

6.2 Supervised methods

Given these input-output pairs, two different supervised learning approaches
(TDNN and kNN) have been evaluated and compared for pitch classification.

4 Given the audio bit depth of the generated audio files (16 bits), the maximum amplitude

in the magnitude spectrogram is 20 log(216 ) = 96 dB.

103
6. MULTIPLE PITCH ESTIMATION USING SUPERVISED LEARNING

pitch

time

...

G!0 F8

Figure 6.1: Binary digital piano-roll coding in each row the active pitches at
each time frame when the spectrogram is computed.

6.2.1 Time-delay neural networks

A time-delay neural network trained with the standard backpropagation
algorithm from Rumelhart et al. (1986) has been used. The scheme of the
network architecture can be seen in Fig. 6.2.
The input data are normalized into the interval [−1, +1], being the −1 value
assigned to the maximum attenuation (ξ dB), and +1 the value assigned to the
attenuation of 0 dB, through this simple equation:

1  
b̄i [t] = (ξ + bi [t]) − 1 6.1 
ξ/2
This way, the input data bi [t] ∈ [ξ, 0] are mapped into b̄i [t] ∈ [−1, +1] for
the network input. Each of these values, which corresponds to one spectral
component, is provided to a neuron at the input layer. The adjacent frames
provide the short-context information. For each frame considered, B new input
units are added to the network, being the total number of input neurons B(n +
m + 1).
The network output layer is composed of B = 94 neurons, one for each
possible pitch. The output is coded in such a way that an activation value of
yk [t] = 1 for a particular unit k means that the k-th pitch is active at that
frame, whereas yk [t] = 0 means that the pitch is not active.
The TDNN has been implemented with bias5 and without momentum6 . The
selected transfer function f (x) is a standard sigmoid (see Fig. 6.3):

2  
f (x) = −1 6.2 
1 + e−x
5 A bias neuron lies in one layer, is connected to all the neurons in the next layer but none

in the previous layer, and it always emits 1.

6 Momentum, based on the notion from physics that moving objects tend to keep moving

unless acted upon by outside forces, allows the network to learn more quickly when there exist
plateaus in the error surface (Duda et al., 2000).

104
 6.2. SUPERVISED METHODS

k[t]= 0, ... , b -1
ν(ti , k) ; ν
.....

................ h[t]
h(t i)

..... ..... ..... ..... .....

S(f
b̄[tk ,−ti-m
m]) S(fb̄[t]
k , ti ) S(f k,+
b̄[t ti+n
n] )
k = 0, ... , b -1 k = 0, ... , b -1 k = 0, ... , b -1

Figure 6.2: TDNN architecture and data supplied during training. The arrows
represent full connection between layers.

11

0.5
0.5

00

-0.5
-0.5

-1-1
-10 -5
-10 -5 00 55 10
10

Figure 6.3: Sigmoid transfer function.

After performing the transfer function, the output values for the neurons are
within the range yk [t] ∈ [−1, +1]. A pitch is detected when yk [t] > α. Therefore,
the activation threshold α controls the sensitivity of the network (the lower is
α, the more likely a pitch is activated).

6.2.2 Nearest neighbors

In the kNN method, the vectors ν[t] are the prototype labels. As previously
discussed in Sec. 2.4.2, in the recognition stage the standard kNN algorithm
can not generalize and find new prototypes not seen in the training stage. A
simple extension of the kNN method has been proposed to mitigate this effect.
In the recognition stage the k nearest neighbors are identified at the target
frame t, and an activation function Ap [t] is obtained for each pitch p:

k
X  
Ap [t] = νp(i) [t] 6.3 
i=0

105
6. MULTIPLE PITCH ESTIMATION USING SUPERVISED LEARNING

(i)
being νp [t] ∈ {0, 1} the corresponding pitch p (not present/present) in the
prototype ν (i) [t]. Then, a low level threshold ζ is established as a fraction of
k, and only the pitches which accomplish Ap [t] ≥ ζ in the neighboring are
considered to be active at the frame t. This way, the method can infer new
prototypes that are not present in the training stage.
Instead of using the number of pitch occurrences as an activation criterion,
additional experiments have been done using weighted distances, summing the
multiplicative inverse of the Eucledian distance di [t] for each neighbor i to
increase the importance of the pitches that are close to the test sample:

k
X 
1
  
A0p [t] = νp(i) [t] 6.4 
i=0
di [t] + 1

A third activation function has been proposed, taking into account the
normalized distances:

k
X 
1

di [t]
  
A00p [t] = νp(i) [t] 1− P 6.5 
i=0
k−1 ∀i di [t]

In all these cases, if the activation function obtains a value greater than ζ,
then the pitch p is added to the prototype yielded at the target frame t.

6.3 Evaluation
A data set of MIDI sequences were utilized for the evaluation of the proposed
methods, obtaining input/output pairs from the MIDI files and the synthesized
audio. Then, 4-folded cross-validation experiments were performed, making four
subexperiments dividing the data set into four parts (3/4 for training and 1/4
for test). The presented results were obtained by averaging the subexperiments
carried out on each data subset. The accuracy of the method is evaluated at
frame-by-frame and note levels.
The frame (or event) level accuracy is the standard measure for multiple
pitch estimation described in Eq. 3.6. A relaxed novel metric has been proposed
to evaluate the system at note level. Notes are defined as series of consecutive
event detections along time. A false positive note is detected when an isolated
series of consecutive false positive events is found. A false negative note is
defined as a sequence of isolated false negative events, and any other sequence
of consecutive event detections is considered as a successfully detected note.
Eq. 3.6 is also used for note level accuracy, considering false positive, false
negative and correctly detected notes.

106
 6.3. EVALUATION

6.3.1 Generation of the training and test sets

The evaluation was done using musical pieces generated with synthetic instru-
ments with near-constant temporal envelopes. The limitations for acoustical
acquisition from real instruments played by musicians and the need of an exact
timing of the ground-truth pitches have conditioned the decision for constructing
these sounds using virtual synthesis models.
Polyphonic tracks of MIDI files (around 25 minutes of music) were impro-
vised by the author and synthesized using different waveshapes, attempting to
have a variety of styles and pitch combinations in the training set. In total,
2, 377 different chords were present in the data set with an average polyphony
of 3 simultaneous sounds. The selected timbres are described next.

Sinusoidal waveshape
This is the simplest periodic wave. Almost all the spectral energy is concentrated
in the f0 component.

Figure 6.4: Sinusoidal waveform and spectrum in dB using a Hanning window.

Sawtooth waveshape
This sound contains all the harmonics with amplitudes proportional to 1/h,
being h the number of harmonic. Only the first H = 10 harmonics were used
to generate this sound.

Figure 6.5: Sawtooth waveform and spectrum in dB using a Hanning window.

Clarinet waveshape
The clarinet sound is generated using a physical model of a clarinet with the
wgclar Csound opcode, which produces good imitating synthesis.

107
6. MULTIPLE PITCH ESTIMATION USING SUPERVISED LEARNING

Figure 6.6: Clarinet waveform and spectrum in dB using a Hanning window.

Hammond organ waveshape

An electrophone timbre corresponding to a Hammond organ7 was selected.

This instrument produces sound through a mechanism based on electromagnetic
induction. In this work, the mechanism has also been simulated using CSound.

Figure 6.7: Hammond waveform and spectrum in dB using a Hanning window.

6.4 Results using time-delay neural networks

6.4.1 Neural network parametrization
Some parameters are free in any neural network. The ones of special interest
in the proposed method are the number of input frames (n + m + 1), and
the activation threshold α. The computational complexity depends on the
number of input frames and, if this value is high, there can be some spectral
frames merging different pitch combinations, which can difficult the training
and recognition processes. The activation threshold α controls the sensitivity
of the network.
These are the most relevant parameters, whereas others concerning the
training, like weight initialization, number of hidden neurons, etc. have shown
to be less important. Different experiments have been carried out varying these
parameters and the results did not vary importantly.
The detailed parametrization results are extensively described in (Pertusa,
2003). After some initial tests, a number of hidden neurons of 100 proved
7 An electronic organ created by L. Hammond in 1934.

108
 6.4. RESULTS USING TIME-DELAY NEURAL NETWORKS

sine sawtooth clarinet Hammond

events 0.94 ± 0.02 0.92 ± 0.02 0.92 ± 0.02 0.91 ± 0.02
notes 0.95 ± 0.02 0.92 ± 0.02 0.92 ± 0.02 0.92 ± 0.02

Table 6.1: Frame-by-frame and note detection accuracy using TDNN.

to be a good choice, with an initial learning rate8 of 0.001. The algorithm

converges when the success rate does not improve during a maximum number
of iterations9 .
According to the size of the input context, the best results were obtained
with m = 1 and n = 0, i.e., considering only one previous frame. Anyway,
the accuracy was not much higher than that obtained with one window at each
side, or even 2+1 or 1+2. The detection was consistently worse with 2+2 and
larger contexts. However, the success rate was clearly worse when no context
was considered, about 20% lower than with any of the other non-zero contexts
tested (see (Pertusa, 2003) for details). Values of α ∈ [−0.8, −0.7] showed to be
adequate.

6.4.2 Recognition results

4-folded cross-validation experiments were performed to assess the capability of
the TDNN for this task with the parameters that showed consistency for the
different waveshapes.
Using this data set, the training algorithm converges relatively fast (in tens
of epochs), and each epoch takes about 5 seconds in a 2 GHz iMac.
The Table 6.1 shows the accuracy and dispersion10 of note and event
detection for the timbres described in Sec. 6.3.1. As expected, the sinusoidal
waveshape provided the best results (around 0.94 for events and 0.95 for notes).
Most of the event detection errors were false negatives at the note boundaries,
and the majority of the errors in note detection corresponded to false negative
notes shorter than 100 ms.
Using the sawtooth timbre, the success rates are lower due to the higher
harmonic content. Around 0.92 for events and also for notes were obtained.
Again, most of the false negatives occurred in very short notes.
For the clarinet and the Hammond organ the results were comparable to
those obtained for the pure synthetic waveshapes, giving values around 0.92
for notes and ranging from 0.90 to 0.92 for events. The results obtained with
8 An adequate learning rate must be selected to ensure that the weights converge to a

response that is neither too specific nor too general.

9 This parameter has been set to 15 iterations.
10 Dispersion is calculated as the difference between the maximum and the minimum

accuracy from the 4 subexperiments divided by 4.

109
6. MULTIPLE PITCH ESTIMATION USING SUPERVISED LEARNING

0.8

Pitch detection accuracy

0.6

0.4

0.2

0
C1 C2 C3 C4 C5 C6 C7 C8
Pitch

Figure 6.8: TDNN recognition accuracy as a function of pitch.

the clarinet and the Hammond suggest that the methodology can be applied to
other instruments characterized by a nearly stable amplitude envelope.
The errors have been analyzed considering note length, pitch, and number
of training samples. Errors produced by notes shorter than 100 ms represent
the 31% of the total amount of errors. With a time resolution ∆t = 46 ms,
these notes extend along one or two frames. Since most of the false negatives
occur at the beginning and end of the notes, these very short notes, which are
not usual in real music, are sometimes missed.
As shown in Fig. 6.8, most of pitch errors correspond to very high (higher
than C7 ) and very low (lower than C3 ) pitches which are very unfrequent in real
music, whereas the method has a very high success rate in the central range of
pitches. This effect is partially related with the amount of musical pitches in
the training set, which is composed of musical data. The most frequent musical
pitches are those at valid central frequencies. There exists a clear correlation of
recognition success for a given pitch to the amount of events in the training set
for that pitch. In Fig. 6.9, each dot represents a single pitch. Abcises represent
the amount of data for that pitch in the training set, whereas ordinates represent
the recognition accuracy. An exponential curve has been adjusted to the data
showing the clear non linear correlation between the amount of training data
and the performance.
Another reason to explain these errors is that lowest pitches are harder to
detect due to the higher frequency precision required, and highest pitches have
less harmonics below Nyquist. Moreover, the harmonics of highest pitches can

110
 6.4. RESULTS USING TIME-DELAY NEURAL NETWORKS

0.8
Pitch detection accuracy

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7
Percentage of pitch events in the training set

Figure 6.9: TDNN correlation between recognition rates for each pitch and the
amount of events in the training set for that pitch.

also produce aliasing when they are synthesized. Anyway, most of the wrong
estimates correspond to very unusual notes that were artificially introduced in
the data set to spread out the pitch range, and which are not common in real
music.
A graphical example of the detection is shown in Fig. 6.10. This musical
excerpt was neither in the training set nor in the recognition set. It was
synthesized using the clarinet timbre11 , and with a fast tempo (120 bpm). In
this example, the event detection accuracy was 0.94, and most of the errors were
produced in the note onsets or offsets. Only 3 very short false positive notes
were detected.

6.4.3 Changing waveshapes for detection

The results for the evaluated waveshapes were similar, showing that the
performance does not critically depends on the selected timbre, at least for
instruments with a fixed spectral profile. To assess how specific the network
weights are for the different timbres considered, musical pieces generated with
a given timbre were presented to a network trained with a different instrument.
The event and note detection results are displayed in tables 6.2 and 6.3,
respectively.
11 Clarinet training weights were also used for recognition.

111
112
C5

C4

C3

C5

C4

C3

0.00 0.44 2.37 3.76 5.15 6.55 sec.

6. MULTIPLE PITCH ESTIMATION USING SUPERVISED LEARNING

Figure 6.10: Temporal evolution of the note detection for a given melody using the clarinet timbre. Top: the original score;
center: the melody as displayed in a sequencer piano-roll; down: the piano-roll obtained from the network output compared
with the original piano-roll. Notation: ‘o’: successfully detected events, ‘+’: false positives, and ‘-’: false negatives.
 6.5. RESULTS USING K NEAREST NEIGHBORS

Events sine sawtooth clarinet Hammond

sine 0.94 ± 0.02 0.57 ± 0.03 0.70 ± 0.04 0.30 ± 0.02
sawtooth 0.48 ± 0.02 0.92 ± 0.02 0.69 ± 0.02 0.26 ± 0.03
clarinet 0.46 ± 0.03 0.67 ± 0.03 0.92 ± 0.02 0.34 ± 0.02
Hammond 0.083 ± 0.014 0.169 ± 0.009 0.144 ± 0.007 0.91 ± 0.02

Table 6.2: Frame level cross-detection results using TDNN. Rows correspond
to training timbres and columns to test timbres.

Notes sine sawtooth clarinet Hammond

sine 0.95 ± 0.02 0.46 ± 0.02 0.61 ± 0.06 0.27 ± 0.02
sawtooth 0.51 ± 0.03 0.92 ± 0.02 0.65 ± 0.01 0.29 ± 0.02
clarinet 0.57 ± 0.04 0.56 ± 0.02 0.92 ± 0.02 0.337 ± 0.008
Hammond 0.089 ± 0.014 0.164 ± 0.009 0.140 ± 0.002 0.92 ± 0.02

Table 6.3: Note level cross-detection results using TDNN. Rows correspond to
training timbres and columns to test timbres.

The accuracy ranges from 0.089 to 0.65 for pitch recognition of sounds
which are different from those used to train the TDNN, showing the network
specialization. The cross-detection value could be an indication of the similarity
between two timbres, but this assumption needs of further in-deep study.

6.5 Results using k nearest neighbors

Initial experiments were done using only 1NN, i.e., classifying the sample with
the closest prototype label. Then, different values for k and ζ have been
evaluated using the three kNN activation functions previously described.
To directly compare the kNN results with the TDNN, the spectral informa-
tion of the previous frame (m = 1, n = 0) was also added to the input.
The evolution of the event detection accuracy using the sinusoidal waveshape
with the different activation functions is shown in Figs. 6.11, 6.12, and 6.13.
These figures are representative of the system behavior with the different
timbres.
In the first activation function Ap , values of ζ have been set as fractions of
k, to make them relative to the number of neighbors. The same occurs with
A0p , but in this case different fractions have been considered, as ζ depends on
the distances, instead of the pitch counts. For the normalized function A00p , the
threshold values vary from 0 to 1.
The experiments using these thresholds and changing the number of
neighbors have been done for all the timbres, obtaining the event and note
accuracy for each of them. In the case of Ap , values of ζ = k/2 and ζ = 2k/3

113
6. MULTIPLE PITCH ESTIMATION USING SUPERVISED LEARNING

1
k/6
k/5
0.9 ζ k/4
k/3
k/2
2k/3
0.8

0.7

0.6

Acc 0.5

0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600 700 800 900 1000

Figure 6.11: Event detection accuracy using Ap for the sinusoidal timbre with
respect to k and ζ.

have provided the best results for event and note detection (see Tabs. 6.4 and
6.5). When k becomes large, the accuracy decreases, and good values for k
are relatively small (from 20 to 50). The behavior is similar for all the tested
timbres. In most cases, Ap obtains a significantly higher accuracy than when
using only one nearest neighbor.
No significant differences were found comparing the best results for A0p and
Ap . However, when using A0p , the number of neighbors does not affect much the
results (see 6.12). The highest accuracy was obtained with ζ ∈ {k/200, k/300}.
The best results for most timbres were obtained using A00 p with k = 20.
Tabs. 6.4 and 6.5 show the success rate for events and notes using k = 20,
which is the best k value among those tested for most timbres (except from the
sinusoidal waveshape, where k = 50 yielded a slightly higher accuracy). It can
be seen that A00p obtains the highest accuracy for most timbres. Anyway, the
best results are significantly worse than those obtained using the TDNN.

6.6 Conclusions
In this chapter, different supervised learning approaches for multiple pitch
estimation have been presented. The input/output pairs have been generated
by sequencing a set of MIDI files and synthesizing them using CSound. The
magnitude STFT apportioned through one semitone filter-bank is used as input

114
 6.6. CONCLUSIONS

1
k/1000
k/300
0.9 ζ k/200
k/100
k/10
k
0.8

0.7

0.6

Acc 0.5

0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600 700 800 900 1000

Figure 6.12: Event detection accuracy using A0p for the sinusoidal timbre with
respect to k and ζ.

1
0.1

0.9
ζ 0.2
0.3
0.4
0.5
0.6
0.8 0.7
0.8
0.9
0.7

0.6

Acc 0.5

0.4

0.3

0.2

0.1

0
0 100 200 300 400 500 600 700 800 900 1000

Figure 6.13: Event detection accuracy using A00p for the sinusoidal timbre with
respect to k and ζ.

115
6. MULTIPLE PITCH ESTIMATION USING SUPERVISED LEARNING

Events k ζ sinusoidal sawtooth clarinet Hammond

1-NN 0.67 ± 0.02 0.62 ± 0.02 0.50 ± 0.02 0.604 ± 0.011
20 k/6 0.74 ± 0.02 0.644 ± 0.011 0.49 ± 0.02 0.637 ± 0.014
20 k/5 0.74 ± 0.02 0.653 ± 0.010 0.48 ± 0.02 0.638 ± 0.011
Ap 20 k/4 0.73 ± 0.02 0.654 ± 0.007 0.50 ± 0.02 0.643 ± 0.010
20 k/3 0.722 ± 0.014 0.653 ± 0.007 0.49 ± 0.02 0.641 ± 0.010
20 k/2 0.668 ± 0.012 0.633 ± 0.011 0.49 ± 0.02 0.612 ± 0.014
20 2k/3 0.62 ± 0.02 0.61 ± 0.02 0.48 ± 0.02 0.58 ± 0.02
20 k/1000 0.61 ± 0.03 0.54 ± 0.03 0.457 ± 0.014 0.55 ± 0.02
20 k/300 0.70 ± 0.02 0.616 ± 0.014 0.49 ± 0.02 0.62 ± 0.02
A0p 20 k/200 0.72 ± 0.02 0.637 ± 0.006 0.49 ± 0.02 0.632 ± 0.010
20 k/100 0.67 ± 0.02 0.633 ± 0.009 0.47 ± 0.03 0.60 ± 0.02
20 k/10 0.22 ± 0.03 0.19 ± 0.03 0.17 ± 0.03 0.16 ± 0.03
20 k 0.21 ± 0.03 0.09 ± 0.03 0.04 ± 0.02 0.08 ± 0.02
20 0.1 0.72 ± 0.03 0.62 ± 0.02 0.48 ± 0.02 0.62 ± 0.02
20 0.2 0.74 ± 0.02 0.651 ± 0.011 0.49 ± 0.02 0.644 ± 0.013
20 0.3 0.727 ± 0.013 0.655 ± 0.007 0.50 ± 0.02 0.643 ± 0.010
20 0.4 0.712 ± 0.013 0.654 ± 0.007 0.49 ± 0.02 0.638 ± 0.007
A00
p 20 0.5 0.681 ± 0.012 0.637 ± 0.011 0.49 ± 0.02 0.623 ± 0.011
20 0.6 0.648 ± 0.012 0.623 ± 0.013 0.49 ± 0.02 0.60 ± 0.02
20 0.7 0.62 ± 0.03 0.60 ± 0.02 0.48 ± 0.02 0.58 ± 0.02
20 0.8 0.58 ± 0.02 0.58 ± 0.02 0.47 ± 0.03 0.54 ± 0.03
20 0.9 0.51 ± 0.03 0.53 ± 0.02 0.44 ± 0.03 0.49 ± 0.03

Table 6.4: Event accuracy using k = 20 for each activation function.

Notes k ζ sinusoidal sawtooth clarinet Hammond

1-NN 0.65 ± 0.02 0.51 ± 0.02 0.46 ± 0.04 0.530 ± 0.013
20 k/6 0.71 ± 0.02 0.55 ± 0.03 0.44 ± 0.04 0.55 ± 0.02
20 k/5 0.722 ± 0.009 0.57 ± 0.03 0.45 ± 0.04 0.57 ± 0.02
Ap 20 k/4 0.728 ± 0.008 0.59 ± 0.03 0.47 ± 0.04 0.59 ± 0.02
20 k/3 0.728 ± 0.009 0.60 ± 0.03 0.48 ± 0.04 0.60 ± 0.02
20 k/2 0.705 ± 0.008 0.614 ± 0.013 0.49 ± 0.04 0.617 ± 0.010
20 2k/3 0.675 ± 0.008 0.612 ± 0.010 0.49 ± 0.05 0.613 ± 0.007
20 k/1000 0.38 ± 0.03 0.29 ± 0.02 0.41 ± 0.03 0.33 ± 0.03
20 k/300 0.61 ± 0.03 0.46 ± 0.03 0.42 ± 0.03 0.50 ± 0.03
A0p 20 k/200 0.688 ± 0.008 0.52 ± 0.03 0.45 ± 0.03 0.57 ± 0.03
20 k/100 0.70 ± 0.02 0.596 ± 0.012 0.45 ± 0.03 0.605 ± 0.009
20 k/10 0.35 ± 0.09 0.30 ± 0.08 0.25 ± 0.08 0.26 ± 0.05
20 k 0.33 ± 0.09 0.18 ± 0.06 0.09 ± 0.04 0.15 ± 0.05
20 0.1 0.62 ± 0.03 0.46 ± 0.03 0.39 ± 0.05 0.47 ± 0.03
20 0.2 0.720 ± 0.010 0.56 ± 0.03 0.45 ± 0.04 0.56 ± 0.02
20 0.3 0.731 ± 0.008 0.59 ± 0.03 0.48 ± 0.04 0.59 ± 0.02
20 0.4 0.733 ± 0.008 0.61 ± 0.02 0.48 ± 0.04 0.62 ± 0.02
A00
p 20 0.5 0.717 ± 0.007 0.611 ± 0.013 0.49 ± 0.04 0.622 ± 0.013
20 0.6 0.697 ± 0.007 0.618 ± 0.010 0.49 ± 0.05 0.620 ± 0.008
20 0.7 0.671 ± 0.011 0.614 ± 0.012 0.48 ± 0.05 0.617 ± 0.008
20 0.8 0.65 ± 0.02 0.600 ± 0.012 0.47 ± 0.05 0.605 ± 0.011
20 0.9 0.60 ± 0.03 0.57 ± 0.02 0.45 ± 0.06 0.57 ± 0.03

Table 6.5: Note accuracy using k = 20 for each activation function.

116
 6.6. CONCLUSIONS

data, whereas the outputs are the ground-truth MIDI pitches. Two different
supervised learning methods (TDNN and kNN) have been used and compared
for this task using simple stationary sounds and taking into account adjacent
spectral frames.
The TDNN performed far better than the kNN, probably due to the huge
space of possible pitch combinations. The results suggest that the neural
network can learn a pattern for a given timbre, and it can find it in complex
mixtures, even in the presence of beating or harmonic overlap. The success
rate was similar in average for the different timbres tested, independently of the
complexity of the pattern, which is one of the points in favour of this method.
The performance using the nearest neighbors is clearly worse than the TDNN
approach. Different alternatives were proposed to generalize in some way the
prototypes matched using the kNN technique to obtain new classes (pitch
combinations) not seen in the training stage. However, these modifications
did not improve significantly the accuracy. An interesting conclusion from this
comparison is that kNN techniques are not a good choice for classification when
there exists many different prototype labels, as in this particular task.
Respect to the TDNN method, errors are concentrated in very low/high
frequencies, probably due to the sparse presence of these pitches in the training
set. This fact suggests that increasing the size and variety of the training set, the
accuracy could be improved. In the temporal dimension, most of the errors are
produced in the note boundaries, which are not very relevant from a perceptual
point of view. This is probably caused by the window length, which can cover
transitions between different pitch combinations. When the test waveshape
was different from that used to train the net, the recognition rate decreased
significantly, showing the high specialization of the network.
The main conclusions are that a TDNN approach can estimate accurately the
pitches in simple waveforms, and the compact input using a one semitone filter-
bank is representative of the spectral information for harmonic pitch estimation.
Future work include to test the feasibility for this approach for real mixtures
of sounds with varying temporal envelopes, but this requires of a large labeled
data set for training, and it is difficult to get musical audio pieces perfectly
synchronized with the ground-truth pitches. However, this is a promising
method that should be deeply investigated with real data.
It also seems reasonable to provide the algorithm with a first timbre
recognition stage, at least at instrument family level. This way, different weight
sets could be loaded in the net according to the decision taken by the timbre
recognition algorithm before starting the pitch estimation.

117
 Multiple fundamental frequency
7
estimation using signal processing
methods

Supervised learning methods require synchronized audio and symbolic data to

be trained. These methods rely on the training set, and for this reason most of
them need a-priori information about the timbre to be transcribed. Probably,
it could be possible to generalize and correctly find the pitches in real audio
files when they are trained using a large data set, but they still depend on the
training data.
In this chapter, three multiple f0 estimation methods are described. These
heuristics approaches purely rely on signal processing methods, avoiding the use
of a training set. The first of them is a simple and efficient iterative cancellation
approach, mainly intended for percussive strings sounds1 . The method was
integrated into a more complex genre classification system published in (Lidy
et al., 2007) to get a basic estimation with a very low computational cost.
Besides the iterative cancellation approach, two novel joint estimation
methods have been proposed. As discussed in Sec. 4.3, joint estimation methods
can model the source interactions better than iterative cancellation approaches.
However, they tend to have higher computational costs due to the evaluation
of many possible source combinations. Contrary to most joint estimation
methods, the proposed techniques have been designed to be very efficient. They
were evaluated and compared to other works in MIREX (2007) and MIREX
(2008) multiple f0 estimation and tracking contests, yielding competitive results
with very efficient runtimes. The first joint estimation approach is detailed
in (Pertusa and Iñesta, 2007, 2008a), and the second in (Pertusa and Iñesta,
2008b).
1 Here, the term percussive string instrument sound is used to refer to the sounds of plucked

and struck string instruments such as piano, guitar and pizzicato violin.

119
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

The methods described in this chapter are implemented in C++, and they
can be compiled and executed from the command line in Linux and Mac
OSX. Two standard C++ libraries have been used, for loading the audio files
(libsndfile2 ), and computing the Fourier transforms (FFTW3 from Frigo and
Johnson (2005)). The rest of code, including the generation of MIDI files, has
been implemented by the author.

7.1 Iterative cancellation method

A simple and efficient iterative cancellation method is described in this section.
The general scheme of the system is shown in Fig. 7.1. In the preprocessing
stage, the STFT is computed, and the sinusoidal components are extracted using
a sinusoidal likeness measure (SLM). The onsets are detected using the method
described in Chapter 5. Then, only the frames that are after each detected onset
are analyzed to yield the active fundamental frequencies in the interval between
two consecutive onsets.
At each analyzed frame, a set of f0 candidates are selected from the sinusoidal
spectral peaks. The candidates are evaluated in ascending frequency order. For
each candidate, the partials are first searched, and a fixed harmonic pattern is
used to subtract the candidate spectral components from the mixture using
iterative cancellation. Finally, a postprocessing stage is done to refine the
estimate, removing pitches with a very low absolute or relative intensity.
The proposed iterative cancellation method is mainly intended for piano
transcription, as the selected harmonic pattern is based on an approximation to
the spectra of most percussive string sounds.

7.1.1 Preprocessing
In the preprocessing stage, the magnitude spectrogram is obtained performing
the STFT with a 93 ms Hanning-windowed frame and a 46 ms hop size. This
window size may seem long for typical signal processing algorithms, but for chord
identification pitch margin is wide (Klapuri, 2003a), and this is also the frame
length used in many previous methods for multiple f0 estimation, like (Klapuri,
2006b; Ryynänen and Klapuri, 2005). Using these spectral parameters, the
temporal resolution achieved is 46 ms. Zero padding has been used, multiplying
the original size of the window by a factor z to complete it with zeroes before
computing the STFT. With this technique, the frequency of the lower pitches
can be more precisely estimated.
Then, at each frame, a sinusoidal likeness measure (SLM) is calculated to
identify the spectral peaks that correspond to sinusoids, allowing to discard the
2 http://www.mega-nerd.com/libsndfile/

120
 7.1. ITERATIVE CANCELLATION METHOD

Waveform

STFT 1/12 octave filterbank

Partial selection (SLM) Onset detection

Candidate selection

Iterative cancellation

Postprocessing

MIDI pitches

Figure 7.1: Scheme of the proposed iterative cancellation method.

spurious non-sinusoidal components. SLM techniques have been successfully

applied to the analysis and synthesis of speech (Griffin and Lim, 1985) and
musical sounds (Doval, 1994, Rodet, 1997, Virtanen, 2000).
The SLM from Rodet (1997) provides a measure vΩ of the likeness between
a pure sinusoid and a partial candidate with frequency Ω. As pointed out by
Virtanen (2000), basically the idea is to calculate the cross-correlation between
the short-time spectrum of the signal and the spectrum resulting from an ideal
sinusoid, scaling the result by the overall spectral shape.
Being H(ω) the DFT of the analysis window3 , and X(ωk ) the amplitude
corresponding to the frequency ωk , the cross-correlation function Γ of the
complex signals H and X within a bandwidth4 W can be calculated as:
X  
Γ(ω) = H(ω − ωk )X(ωk ) 7.1 
k,|ω−ωk |<W

Each maximum |Γ(Ω)| indicates a sinusoidal partial candidate at frequency

Ω. Defining norms for H(ωk ) and X(ωk ) for a frequency Ω:

3 In the proposed method, a Hanning function.

4 SLM assumes that there are not two sinusoids closer than W Hz.

121
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

60

50

40

30

20

10

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000

Figure 7.2: Example magnitude spectrum (top) and SLM (bottom) for two
sounds in an octave relation (92.5 and 370 Hz) using W = 50 Hz. The
fundamental frequencies are indicated with arrows.

X  
|H|2Ω = |H(Ω − ωk )|2 7.2 
k,|Ω−ωk |<W

X  
|X|2Ω = |X(ωk )|2 7.3 
k,|Ω−ωk |<W

the SLM at a frequency Ω can be computed as:

|Γ(Ω)|  
vΩ = 7.4 
|H|Ω |X|Ω
An efficient implementation of vΩ has been chosen using the method
proposed by Virtanen (2000). The cross-correlation of the frequency domain
is performed through a multiplication of the time-domain signals, and Γ(ω) is
calculated using the DFT for x[n] windowed twice with the Hanning function.
The calculation of |X|2Ω can be implemented with a IIR filter which has only
two non-zero coefficients: one delay takes a cumulative sum of the signal and
the other subtracts the values at the end of the window.
After this process, a sinusoidal likeness function (see Fig. 7.2) is obtained
at each frame. The harmonic peak selection is done as follows: if there is a
peak in the SLM function which value is vΩ > τ , being τ a constant threshold,

122
 7.1. ITERATIVE CANCELLATION METHOD

then the original spectral component at the same frequency Ω, with its original
amplitude, is added to the harmonics list. The spectral components that
do not satisfy the previous condition are discarded. Therefore, the mid-level
representation of the proposed method consists on a sparse vector containing
only certain values of the original spectrum (those ideally corresponding to
partials). This sparse representation reduces the computational cost with
respect to the analysis of all spectral peaks.

7.1.2 Onset detection

The onsets are detected from the STFT using the method described in
Chapter 5, identifying each frame at ti as onset or not-onset. For efficiency,
only a single frame between two consecutive onsets is analyzed to yield the
fundamental frequencies within that inter-onset interval.
To avoid the analysis of transients at the onset times, the frame chosen to
detect the active notes is to + 1, being to the frame where an onset was detected.
Therefore, only those frames that are 46 ms after a detected onset are analyzed
to estimate the fundamental frequencies in the interval between two adjacent
onsets. The scheme for estimating the pitches only between two consecutive
onsets has also been used in the recent method from Emiya et al. (2008b).

7.1.3 Candidate selection

For each frame to + 1, a selection of f0 candidates is done from the sparse array
of sinusoidal peaks. Therefore, as many other iterative cancellation approaches,
the method assumes that the partial corresponding to the fundamental fre-
quency is present. This assumption is made to improve the efficiency, as it
reduces significantly the number of candidates respect to the analysis of the
whole spectral range.
There are two restrictions for a peak to be a candidate; only candidates
within a given pitch margin [fmin , fmax ] are considered, and the difference
between the candidate frequency and the frequency of the closest pitch in
the equal temperament must be lower than fd Hz5 . This is a constant value
introduced to remove some false candidates at high frequencies.

7.1.4 Iterative cancellation

The candidates are sorted in ascending frequency order, and they are evaluated
using the iterative cancellation method described in Alg. 1.
The partials of each candidate are searched taking inharmonicity into
account. Similarly to (Bello et al., 2002), (Every and Szymanski, 2006), and
5 This technique is not suitable for not tuned instruments.

123
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

Algorithm 1: Iterative cancellation method

Input: Sinusoidal peaks and f0 candidates at a given frame
Output: Estimated fundamental frequencies

residual ← sinusoidal peaks

for each f0 candidate do
intensity ← f0 amplitude
for h = 2 to H do
find partial h in residual
if partial not found then
missingPartials ← missingPartials +1
else
expected ← f0 amplitude · p[h − 1]
obtained ← partial amplitude
if expected > obtained then
Remove partial from the residual
else
residual[partial] ← residual[partial] − expected
end
end
intensity ← intensity + expected
if missingPartials > allowed missing partials then
Discard candidate
end
end
if Candidate is discarded then
Residual is not updated
else
Add candidate to the f0 list
end
end

(Emiya et al., 2008b), the partial frequency deviation model proposed by

Fletcher and Rossing (1988) for piano sounds, and previously described in Eq.
2.17, has been used with β = 0.0004 (Fletcher and Rossing, 1988). A harmonic
h is identified as the closest spectral peak to hf0 within the margin [hf0 ± fh ].
When a sinusoidal peak is identified as a candidate partial, its expected
magnitude6 is calculated according to a fixed harmonic pattern p and then
subtracted from the residual. The chosen harmonic pattern represents the
relative amplitudes of the first H = 8 harmonics with respect to the amplitude
of the f0 with the following weights:

6 i.e., the expected contribution to the mixture.

124
 7.1. ITERATIVE CANCELLATION METHOD

Sinusoidal peaks. A candidate is found

Constant harmonic pattern centered at

the candidate frequency and scaled with
the fundamental frequency amplitude

Residual peaks after cancellation

Figure 7.3: Candidate cancellation example.

 
p = {1, 0.5, 0.4, 0.3, 0.2, 0.1, 0.05, 0.01} 7.5 

Therefore, the harmonic pattern is a vector where each component represents

a harmonic amplitude relative to the f0 amplitude, and the first component is
always 1. The selected pattern is adequate for percussive string instruments,
as it approximates the spectral envelope of a piano sound and has a similar
sawtooth shape to the 1/h model proposed by Klapuri (2006b).

When a partial is found, its expected amplitude is set as the f0 amplitude

multiplied by p[h − 1]. If this value is greater than the actual spectral peak
amplitude, the sinusoidal peak is completely removed. Otherwise, the expected
amplitude is linearly subtracted from the corresponding spectral peak (see
Fig. 7.3), thus assuming the additivity of linear spectrum.

Candidates without a minimum number of found harmonics experimentally

set as H/2 are discarded. The intensity ln of a note is finally calculated as the
sum of the expected harmonic amplitudes.

125
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

7.1.5 Postprocessing
Those candidates with a low absolute or relative intensity are removed. First,
the pitch candidates with a intensity ln < γ are discarded. The maximum
note intensity L = max∀n {ln } at the target frame is calculated to remove the
candidates with ln < ηL, as the sources in the mixture should not have very
important energy differences7 .
Finally, the frequencies of the selected candidates are converted to MIDI
pitches with Eq. 2.21. Using this inter-onset based scheme, there are certain
ambiguous situations that are not produced in a frame by frame analysis. If a
pitch is detected in the current and previous inter-onset interval, then there are
two possibilities: there exists a single note spanning both onsets, or there is a
new note with the same pitch.
To make a simple differentiation between new notes and detections of pitches
that were already sounding in the previous frames, the estimation is done at
frames to + 1, and to − 1. If a detected pitch at frame to + 1 is not detected
at to − 1, then a new note is yielded. Otherwise, the note is considered to be a
continuation of the previous estimate.

7.2 Joint estimation method I

A novel joint evaluation method for multiple f0 estimation is presented in this
section. In contrast to most previous joint approaches, this method has a very
low computational cost keeping a high accuracy.
The overall scheme of the method can be seen in Alg. 2. The system performs
a frame by frame analysis, yielding a set of pitch estimates at each instant. A set
of pitch candidates are first identified from the spectral peaks at each frame8 .
All the possible candidate combinations are generated, and a joint algorithm is
used to find the best combination taking into account the source interactions.
To evaluate a combination, a hypothetical partial sequence (HPS9 ) is built
for each candidate. A candidate score (salience) is calculated taking into account
the sum of its HPS harmonic amplitudes and the smoothness measure of the
spectral envelope. The salience of a combination is computed as the sum of
the squared candidate saliences, and the combination with highest salience is
selected at the target frame.
The method assumes that the spectral envelopes of the analyzed sounds tend
to vary smoothly as a function of frequency. The spectral smoothness principle
7 When this occurs, the notes with lower energy are hardly perceived, being sometimes

masked by the other harmonic components in the mixture.

8 In the proposed joint estimation methods, SLM is not used.
9 HPS is the term proposed by Yeh et al. (2005) to refer a vector containing hypothetical

partial amplitudes (see pag. 65).

126
 7.2. JOINT ESTIMATION METHOD I

Algorithm 2: Joint estimation method I

Input: Spectral peaks at a given frame
Output: Estimated fundamental frequencies

Select candidates from spectral peaks

Generate all possible candidate combinations
for each combination C do
residual ← spectral peaks
for each candidate c ∈ C do
estimate HPS from residual
residual ← residual − HPS amplitudes
evaluate HPS salience
end
combination salience ← sum of squared HPS saliences
end
return combination with max salience

has also been used in different ways in the literature (Klapuri, 2003a, Yeh et al.,
2005, Cañadas-Quesada et al., 2008, and Zhou et al., 2009). The proposed novel
smoothness measure is based on the convolution of the hypothetical harmonic
pattern with a gaussian window.
Given a combination, the HPS of each candidate is calculated considering the
harmonic interactions with the partials of all the candidates in the combination.
The overlapped partials are first identified, and their amplitudes are estimated
by linear interpolation using the non-overlapped harmonic amplitudes.
In contrast with the previous iterative cancellation method, which assumes a
constant harmonic pattern, the proposed joint approach can estimate hypothet-
ical harmonic patterns from the spectral data, evaluating them according to the
properties of harmonic sounds. This approach is suitable for most real harmonic
sounds, in contrast with the iterative method, which assumes a constant pattern
based in percussive string instruments.

7.2.1 Preprocessing
In the preprocessing stage, the STFT is computed using a 93 ms Hanning
windowed frame, with a 9.28 ms hop size. The frame overlap ratio may seem
high from a practical point of view, but it was required to compare the method
with other works in the MIREX (2007) evaluation contest (see Sec. 7.4.3). Like
in the iterative cancellation method, zero padding has been used to get a more
precise estimation of the lower frequencies.
SLM has not been used in the joint estimation approaches. Experimentally,
including SLM in the iterative cancellation algorithm did not improve the results
(see Sec. 7.4.1 for details), so it was removed in the joint estimation methods

127
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

that were implemented subsequently. Instead, a simple peak picking algorithm

was used to extract the hypothetical partials from the magnitude spectrum.
At each frame, the spectral peaks with an amplitude higher than a given
constant threshold µ are selected, discarding the rest of spectral information and
obtaining this way a sparse representation which only contains certain spectral
bins for the following analysis.

7.2.2 Candidate selection

The evaluation of many combinations of candidates increases significantly the

computational cost of the algorithm. To improve the efficiency, the candidates
are first ordered decreasingly by the sum of their harmonic amplitudes and,
at most, only the first F candidates of this ordered list are considered. The
adequate selection of candidates plays an important role in joint estimation
methods, and a good compromise between the accuracy and computational cost
(which depends on the number of candidates) must be chosen.
Like in the iterative cancellation approach, the f0 candidates are selected
from the spectral peaks10 that fulfill a series of requirements. First, a candidate
must be within the range [fmin , fmax ], which corresponds to the pitches of
interest. If a spectral peak amplitude is lower than a threshold ε, then the peak
is discarded as a candidate.
To search for the partials, a constant margin fh ± fr around each harmonic
frequency is considered, allowing slight harmonic deviations. The closest peak
to the center of this margin is set as a found partial. If there are no spectral
peaks within this margin, a missing harmonic is considered.
Like in (Yeh, 2008), the harmonic spectral location and spectral interval
principles (Klapuri, 2004) have been considered, taking inharmonicity into
account. The ideal frequency fh of the first partial is initialized to fh = 2f0 .
The next partials are searched at fh+1 = (fx + f0 ) ± fr , where fx = fi if the
previous harmonic h was found at the frequency fi , or fx = fh if the previous
partial was not found.
In some methods, the candidate selection is done according to harmonicity
criterion (Yeh, 2008), partial beating (Yeh, 2008), or the product of harmonic
amplitudes in the power spectrum (Emiya et al., 2008b). The proposed method
uses the sum of harmonic amplitudes as a score function for candidate selection,
avoiding that finding a harmonic with an amplitude lower than 1 will decrease
the score, as it occurs when the harmonic magnitudes are multiplied.

10 As the candidates are spectral peaks, timbres with missing fundamental are not

considered, like in the iterative cancellation method.

128
 7.2. JOINT ESTIMATION METHOD I

7.2.3 Generation of combinations of candidates

All the possible candidate combinations are calculated and evaluated, and
the combination with highest salience is selected at the target frame. The
combinations consist of different number of pitches. In contrast with other
works, like (Yeh et al., 2005), there is not need for a-priori estimation of the
number of concurrent sounds before detecting the fundamental frequencies, and
the polyphony is implicitly calculated in the f0 estimation stage, selecting the
combination with highest score independently from the number of candidates.
At each frame, a set of combinations {C1 , C2 , . . . , CN } is obtained. For
efficiency, like in the recent approach from Emiya et al. (2008b), only the
combinations with a maximum polyphony P are generated from the F
candidates. The amount of combinations N without repetition can be calculated
as:

P  
X F
P
X F!  
N= = 7.6 
n=1
n n=1
n!(F − n)!

This means that when the maximum polyphony is P = 6 and there are
F = 10 selected candidates, N = 847 combinations are generated. Therefore,
N combinations are evaluated at each frame, and the adequate selection of F
and P is critical for the computational efficiency of the algorithm.

7.2.4 HPS estimation

The candidates of a combination C are ordered by ascending frequency. Then, a
HPS vector pc , consisting on the hypothetical harmonic amplitudes of the first
H harmonics, is estimated for each candidate c ∈ C:
 
pc = {pc,1 , pc,2 , . . . , pc,h , . . . , pc,H } 7.7 
where pc,h is the amplitude of the h harmonic of the candidate c. The partials
are searched as previously described for the candidate selection stage. If a
particular harmonic is not found, then the corresponding value pc,h is set to
zero.
Once the partials of a candidate are identified, the HPS values are estimated
considering hypothetical source interactions. To do it, the harmonics of all the
candidates in the combination are first identified, and they are labeled with the
candidate they belong to (see Fig. 7.4). After the labeling process, there are
harmonics that only belong to one candidate (non-overlapped harmonics), and
harmonics belonging to more than one candidate (overlapped harmonics).
Assuming that the interactions between non-coincident partials (beating)
do not alter significantly the original spectral amplitudes, the non-overlapped

129
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

A Spectral peaks

f
f1 f2
Partial identification

A ff11
ff22

f
f1 f2
Linear subtraction

A A

f f
f1 HPS(f1) HPS(f2)
f2

Figure 7.4: Interpolation example in a combination of two candidates separated

by an octave. The HPS are estimated with the interpolated values.

amplitudes are directly assigned to the HPS. However, the contribution of each
source to an overlapped partial amplitude must be estimated. This can be done
using the amplitudes of non-overlapped neighbor partials (Klapuri, 2003a, Yeh
et al., 2005, Every and Szymanski, 2006), assuming smooth spectral envelopes,
or considering that the amplitude envelopes of different partials are correlated
in time (Woodruff et al., 2008).
In the proposed method, similarly to (Maher, 1990) and (Yeh et al., 2005),
the amplitudes of overlapped partials are estimated by linear interpolation of
the neighboring non-overlapped partials (see Fig. 7.4).
If there are more than two consecutive overlapped partials, then the
interpolation is done the same way with the non-overlapped values. For instance,
if harmonics 2 and 3 are overlapped, then the amplitudes of harmonics 1 and 4
are used to estimate them by linear interpolation.

130
 7.2. JOINT ESTIMATION METHOD I

After the interpolation, the contribution of each partial to the mixture

must be estimated and subtracted before processing the next candidates. This
calculation is done as follows:

• If the interpolated (expected) value is greater than the corresponding

overlapped harmonic amplitude, then pc,h is set as the original harmonic
amplitude, and the spectral peak is completely removed from the residual,
setting it to zero for the candidates that share that partial.
• If the interpolated value is smaller than the corresponding overlapped
harmonic amplitude, then pc,h is set as the interpolated amplitude, and
this value is linearly subtracted for the candidates that share the harmonic
(see Fig. 7.4).
Therefore, the additivity of linear spectrum (cos(φ∆ ) = 0) is assumed to
estimate the amplitude of colliding partials. The residual harmonic amplitudes
after this process are iteratively analyzed for the rest of the candidates in the
combination in ascending frequency order.

7.2.5 Salience of a combination

Once the HPS of all the candidates have been estimated for a given combination,
their saliences are calculated. The salience of a combination is the squared sum
of the saliences of its candidates. A candidate salience is obtained taking into
account the intensity and the smoothness of its HPS.
The intensity l(c) of a candidate c is a measure of the strength of a source,
and it is computed as the sum of the HPS amplitudes:
H
X  
l(c) = pc,h 7.8 
h=1

Like in other works, the method also assumes that a smooth spectral pattern
is more probable than an irregular one. To compute the smoothness σ of a
candidate, the HPS is first normalized dividing the amplitudes by the maximum
harmonic value in the HPS, obtaining p̄. Then, p̄ is low-pass filtered using a
truncated normalized Gaussian window N0,1 , which is convolved with the HPS
to obtain the smoothed version p̃:
 
p̃c = N0,1 ∗ p̄c 7.9 
Only three components were chosen for the Gaussian window N =
{0.21, 0.58, 0.21}, due to the reduced size of the HPS11 .
11 Usually, only the first harmonics contain most of the energy of a harmonic source, therefore

typical values for H are within the margin H ∈ [5, 20].

131
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

1
p
0.8 ~
p
0.6

0.4

0.2

1 2 3 4 5 6 7 8 9 10 h
1
p
0.8
~
p

0.6

0.4

0.2

1 2 3 4 5 6 7 8 9 10 h

Figure 7.5: Spectral smoothness measure example. The normalized HPS vector
p and the smooth version p̃ of two candidates (c1 , c2 ) are shown. Sharpness
values are s(c1 ) = 0.13, and s(c2 ) = 1.23.

Then, as shown in Fig. 7.5, a sharpness measure s(c) is computed by sum-

ming the absolute differences between the smoothed values and the normalized
HPS amplitudes:

H
X  
s(c) = (|p̃c,h − p̄c,h |) 7.10 
h=1

The sharpness s(c) is normalized into s̄(c):

s(c)  
s̄(c) = 7.11 
1 − N0,1 (x̄)
And finally, the smoothness σ(c) ∈ [0, 1] of a HPS is calculated as:

s̄(c)  
σ(c) = 1 − 7.12 
Hc
where Hc is the index of the last harmonic found for the candidate. This
parameter was introduced to prevent that high frequency candidates that have
less partials than those at low frequencies will have higher smoothness. This
way, the smoothness is considered to be more reliable when there are more
partials to estimate it.

132
 7.3. JOINT ESTIMATION METHOD II

Once the smoothness and the intensity of each candidate have been
calculated, the salience S(Ci ) of a combination Ci with C candidates is:

C
X 2
 
S(Ci (t)) = [l(c) · σ κ (c)] 7.13 
c=1

where κ is a factor to control the smoothness weight.

Note that when there are overlapped partials, their amplitudes are estimated
by interpolation, and the HPS smoothness tends to increase. To partially
compensate this effect in the salience function, the candidate saliences are
squared in Eq. 7.13 before summing them. This way, the method assigns a
higher score to a combination of one candidate with a given salience than to
a combination of two candidates with half salience values. The square factor
favors a sparse representation, as it is convenient to explain the mixture with
the minimum number of sources. Experimentally, it was found that this square
factor was very important to improve the success rate of the method (see pag.
147).
The combination with highest salience is selected at the target frame.
Similarly to the iterative cancellation approach, combinations that have at least
one candidate with l(c) < η or l(c) < γL are discarded, being L = max∀c {l(c)}
the maximum intensity of the candidates in the combination.

7.2.6 Postprocessing
After selecting the best combination at each individual frame, a last stage is
applied to remove some local errors taking into account the temporal dimension.
If a pitch was not detected in a target frame but it was found in the previous
and next frames, it is considered to be active in the current frame too, avoiding
some temporal discontinuities. Notes shorter than a minimum duration d are
also removed.
Finally, the sequences of consecutive detected fundamental frequencies are
converted into MIDI pitches. The maximum intensity of the entire song
candidates is used as reference to get the MIDI velocities, linearly mapping
the candidate intensities within the range [0, max∀C,c {l(c)}] into MIDI values
[0, 127].

7.3 Joint estimation method II

In the previous joint estimation method, each frame was independently
analyzed, yielding the combination of fundamental frequencies which maximizes
a salience measure. One of the main limitations of this approach is that the
selected window size (93 ms) is relatively short to perceive the pitches in a

133
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

complex mixture, even for expert musicians. As discussed in Sec. 3.1, context is
very important in music to disambiguate certain situations. The joint estimation
method II is an extension of the previous method, but considering information
about adjacent frames, similarly to the supervised learning method described
in Chapter 6, producing a smoothed detection across time.

7.3.1 Temporal smoothing

In this method, instead of selecting the combination with highest salience at
isolated frames, adjacent frames are also considered to get the salience of each
combination of pitches, performing a temporal smoothing.
In order to merge similar information across time, the frequencies of each
combination C are first converted into pitches using Eq. 2.21, obtaining a pitch
combination C 0 . For instance, the combination Ci = {261 Hz, 416 Hz} is mapped
into Ci0 = {C3 , G]3 }. If there are more than one combination with the same
pitches at a target frame, it is removed, keeping only the combination with the
highest salience value.
Then, at a target frame t, a new smoothed salience function S̃(Ci0 (t)) for a
combination Ci0 is computed using the neighbor frames:

t+K
X  
S̃(Ci0 (t)) = S(Ci0 (j)) 7.14 
j=t−K

This way, the saliences of the combinations with the same pitches than Ci0 in
the K adjacent frames are summed to obtain the salience at the target frame, as
shown in Fig. 7.6. The combination with maximum salience is finally selected
to get the pitches at the target frame t.
 
C 0 (t) = arg max{S̃(Ci0 (t))} 7.15 
i

This new approach increases the robustness of the system in the data set
used for evaluation, and it allows to remove the minimum amplitude ε for a peak
to be a candidate, added in the previous approach to avoid local false positives.
If the selected combination at the target frame does not contain any pitch
(if there is not any candidate or if none of them can be identified as a pitch),
then a rest is yielded without evaluating the combinations in the K adjacent
frames.
This technique smoothes the detection in the temporal dimension. For a
visual example, let’s consider the smoothed intensity of a given candidate c0 as:

t+K
X  
˜l(c0 (t)) = l(c0 (j)) 7.16 
j=t−K

134
 7.3. JOINT ESTIMATION METHOD II

Combinations at t − 1, t and t + 1 using K = 1. Best

combination at frame t using method I is hightlighted

C1! (t − 1) = {C3} S(C1! (t − 1)) =2100

C2! (t − 1) = {G3, E3} S(C2! (t − 1)) =1000

Best combination at frame t using method II
C3! (t − 1) = {E3, G4} S(C3! (t − 1)) =140
C1! (t) ={C3} S̃(C1! (t)) =7400

C1! (t) ={C3, G4} S(C1! (t)) =2000 C2! (t) ={C3, G4} S̃(C2! (t)) =3200

C2! (t) ={C3} S(C2! (t)) =1800 C3! (t) ={G3, E3} S̃(C3! (t)) =1000

C3! (t) ={E3, G4} S(C3! (t)) =200 C4! (t) ={E3, G4} S̃(C4! (t)) =340

C5! (t) ={E3} S̃(C5! (t)) =100

C1! (t + 1) = {C3} S(C1! (t + 1)) =1700

C2! (t + 1) = {C3, G4} S(C2! (t + 1)) =1200

C2! (t + 1) = {E3} S(C3! (t + 1)) =100

Figure 7.6: Example of combinations merged across adjacent frames using

K = 1. The selected combination differs using methods I and II.

135
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

Figure 7.7: Top: Example of detected piano-roll for an oboe melody. Bottom:
Three-dimensional temporal representation of ˜l(c0 (t)) for the candidates of the
winner combination at each frame. In this example, all the pitches were correctly
detected. High temporal smoothness usually indicates good estimates.

When the temporal evolution of the smoothed intensities ˜l(c0 (t)) of the
winner combination candidates is plotted in a three-dimensional representation
(see Figs. 7.7 and 7.8), it can be seen that the correct estimates usually show
smooth temporal curves. An abrupt change (a sudden note onset or offset,
represented by a vertical cut in the smoothed intensities 3D plot) means that the
harmonic components of a given candidate were suddenly assigned to another
candidate in the next frame. Therefore, vertical lines in the plot usually indicate
errors mapping harmonic components with the candidates.

7.3.2 Partial search

Like in the joint estimation method I, a partial is searched at fh+1 = (fx + f0 ) ±
fr , being fx = fi if the previous harmonic h was found at the frequency fi , or
fx = fh otherwise.
However, in the method II, instead of just selecting the closest peak, a
triangular window has been used. This window, centered in fh with a bandwith
2fr and a unity amplitude, has been used to weight the partial magnitudes
within this range (see Fig. 7.9). The spectral peak with maximum weighted

136
 7.3. JOINT ESTIMATION METHOD II

Figure 7.8: Top: Example of detected piano-roll for a mixture of sounds.

False positives are marked with circles and false negatives with horizontal red
bars. Bottom: Three-dimensional temporal representation of ˜l(c0 (t)) for the
candidates of the winner combination at each frame. Note that most errors
occur when there exist vertical transitions which occur when harmonics are
suddenly reassigned to another candidate.

137
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

Selected partial

0
fh − fr fh fh + fr f

Figure 7.9: Partial selection in the joint estimation method II. The selected
peak is the one with the greatest weighted value.

value is selected as a partial. The advantage of this scheme is that low amplitude
peaks are penalized and, besides the harmonic spectral location, intensity is also
considered to identify the most important spectral peaks with partials.

7.3.3 Fundamental frequency tracking

A simple f0 tracking technique has been introduced to increase the temporal
coherence using a weighted directed acyclic graph (wDAG). The idea is to apply
a penalty when there exist abrupt changes in the estimate, favoring this way
smoother temporal detections. This is done by measuring the differences in the
intensities of the candidates between two consecutive frames.
Let G = (V, E, w) be a wDAG, with vertex set V , edge set E, and edge
function w, where w(vi , vj ) is the weight of the edge from the vertex vi to vj .
Each vertex vi ∈ V represents a combination Ci0 . The vertices are organized in
layers (see Fig. 7.10). Each layer Vt contains the M combinations with highest
salience at a time frame t.
The edges connect all the vertices of a layer with all the vertices of the next
layer, in a way that, if (vi , vj ) ∈ E, then vi ∈ Vt and vj ∈ Vt+1 . The weights
w(vi , vj ) between two combinations are computed as follows:

D(vi , vj )  
w(vi , vj ) = 7.17 
S(vj ) + 1
where S(vj ) is the salience of the combination at vertex vj and D(vi , vj ) is a
similarity measure for two combinations vi and vj , corresponding to the sum
of the absolute differences between the intensities of all the candidates in both
combinations:

X X X  
D(vi , vj ) = |˜l(vi,c ) − ˜l(vj,c )| + ˜l(vi,c ) + ˜l(vj,c )
7.18 
∀c∈vi ,vj ∀c∈vi −vj ∀c∈vj −vi

138
 7.3. JOINT ESTIMATION METHOD II

24609.9

18211.3
{C3,C5}
t4n48n72 15991.3
{C3,C4}
t5n48n60
16638
18214.9
155.338

0
56220.1 {C3,C5}
t6n48n72
9770.75
14504.2
13916.6

23555.4 9862.21 {C3,C5}

t7n48n72
0 10731.9
init {C3}
t4n48 14180.2
16066.6 7765.62
0
10186.3 11630.3 {C3}
t8n48
{C3,C5}
t5n48n72
40399.4 939.891
17735.3 436.046
15925.1
14971 10126.8
{C5}
t4n72 {C3}
t6n48
9822.03 7036.34 553.255
{C3,C4}
t9n48n60
18393 11539.2
{C3}
t5n48 {C3}
t7n48 16489.6

53811.1 15074 7710.18

0
{C3,C4}
t6n48n60 7262.77 {C3,G4}
t8n48n60 118287
1828.94
10186.3 8359.9
32558.2
0
{C3,C4}
t7n48n60 {C3}
t9n48 end

7262.77 16014.2
Vi 0
95375.4
{C3,C5}
t8n48n72
32558.2

{C2,G4}
t9n35n60

83984.6

Figure 7.10: wDAG example for M = 3 combinations at each time frame.

Each vertex represents a combination Ci0 . The vertices are organized in columns
which correspond to the Vi layers. Weights have been multiplied by 106 for
visual clarity. The shadowed nodes are the pitch combinations selected at each
frame. In this example, all the pitches were correctly detected.

Using this scheme, the transition weight between two combinations considers
the salience of the target combination and the differences between the candidate
intensities.
Finally, the shortest path12 across the wDAG is found using the Dijkstra
(1959) algorithm13 . The vertices that belong to the shortest path are the winner
combinations yielded at each time frame.

7.3.4 Alternative architectures

When using the joint estimation method II to identify the fundamental
frequencies, there are different new alternative architectures that can be
considered:

1. Frame by frame analysis. All the frames are analyzed to yield the
estimates. This is the basic scheme of the joint estimation methods
previously described.
12 The path which minimizes the weights sum from the starting node to the final state.
13 The boost C++ library, available at http://www.boost.org, was used for this task.

139
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

2. To detect onsets and analyze only one frame between two onsets to yield
the pitches in the inter-onset interval. This scheme, used in the iterative
estimation method, increases the efficiency but with an accuracy cost. The
method relies on the onset detection results, therefore a wrong estimate
in the onset detection stage can affect the results.

3. To detect onsets and merge combinations of those frames that are between
two consecutive onsets, yielding the pitches for the inter-onset interval.
This technique can obtain more reliable results when the onsets are
correctly estimated, as it happens with piano sounds. However, merging
combinations between two frames reduce the number of detected notes,
as only combinations that are present in most of the IOI frames are
considered. Like in the previous scheme, the detection is very sensitive
to false negative onsets.

4. To detect beats and merge combinations with a quantization grid. Once

the beats are estimated14 , a grid split with a given beat divisor 1/q can
be assumed, considering that there are not triplets and that the minimum
note duration is q. For instance, if q = 4, each inter-beat interval can
be split in q sections, each one of a one sixteenth length. Then, the
combinations of the frames that belong to the quantization unit can be
merged to obtain the results at each minimum grid unit. Like in the onset
detection scheme, the success rate of this approach depends on the success
rate of beat estimation.

The implementation of the joint estimation method II allows to run the

algorithm using any of these schemes. The adequate choice of the scheme
depends on the signal to be analyzed. For instance, for percussive timbres, it is
recommended to use the third scheme, as usually onset detection is very reliable
for this kind of sounds. These architectures have been perceptually evaluated
using some example real songs, but rigorous evaluation of these schemes is left
as future work, since an aligned dataset of real musical pieces with symbolic
data is required for this task.
In order to obtain a more readable score, the tempo changes can optionally
be written into the output MIDI file. To do it, the system allows to load a list
of beat times. A tempo T = 60/Tb is reestimated at each beat instant using the
temporal difference Tb between the current and the previous beat. There is not
other metrical information extracted, therefore the bar impulses are sometimes
shifted due to anacrusis15 or incorrect time signature, which is always assumed
to be of 4/4, since this is the most frequently used musical meter.
14 Beats can be estimated with an external beat tracking algorithm like BeatRoot from

Dixon (2006).
15 Like it occurs in Fig. 1.1.

140
 7.4. EVALUATION

7.4 Evaluation
To perform a first evaluation and set up the parameters, initial experiments were
done using a data set of random mixtures. Then, the three proposed approaches
were evaluated and compared with other works for real music transcription in
the MIREX (2007) and MIREX (2008) multiple f0 estimation and tracking
contests.

7.4.1 Parametrization
The parameters of the three proposed methods and their impact on the results
are analyzed in this section. The intention in the parametrization stage is
not to get the parameter values that maximize the accuracy for the test set
used, as the success rate is dependent on these particular data. However, this
stage can help to obtain a reasonable good parameter set of values and to
evaluate the impact of each parameter in the accuracy and the computational
cost. Therefore, the selected parameters are not always those that achieve the
highest accuracy in the test set, but those that obtain a close-to-best accuracy
keeping a low computational cost.
For the parametrization stage, a database of random pitch combinations has
been used. This database was generated using mixtures of musical instrument
samples with fundamental frequencies ranging between 40 and 2100 Hz. The
samples are the same used in the evaluation of the Klapuri (2006b) method.
The data set consists on 4000 mixtures with polyphony16 1, 2, 4, and 6. The
2842 audio samples from 32 musical instruments used to generate the mixtures
are from the McGill University master samples collection17 , the University of
Iowa18 , IRCAM studio online19 , and recordings of an acoustic guitar. In order
to respect the copyright restrictions, only the first 185 ms of each mixture20
were used for evaluation.
It is important to note that the data set only contains isolated pitch
combinations, therefore the evaluation of the parameters that have a temporal
dimension (like the minimum note duration) could not be evaluated using this
database. The test set is intended for evaluation of multiple f0 estimation at
single frames, therefore f0 tracking from joint estimation method II could not
be evaluated with these data.
To evaluate the parameters in the iterative cancellation method and in
the joint estimation method I, only one frame which is 43 ms apart from the
16 There are 1000 mixtures for each polyphony.
17 http://www.music.mcgill.ca/resources/mums/html/index.htm
18 http://theremin.music.uiowa.edu/MIS.html
19 http://forumnet.ircam.fr/402.html?&L=1
20 Thanks to A. Klapuri for providing this reduced data set for evaluation.

141
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

Stage Parameter Symbol Value

SLM bandwidth W 50 Hz
Preprocessing SLM threshold τ 0.1
Zero padding factor z 8
Candidate selection f0 range [fmin , fmax ] [38, 2100] Hz
closest pitch distance fd 3 Hz
Postprocessing Absolute intensity threshold γ 5
Relative intensity threshold η 0.1

Table 7.1: Iterative cancellation method parameters. Shadowed parameters

(W and τ ) were removed after the initial experiments, as SLM did not improve
the accuracy.

beginning of the mixture has been selected. For the joint estimation method II,
which requires more frames for merging combinations, all the frames (5) have
been used to select the best combination in the mixture.
The accuracy metric (Eq. 3.6) has been chosen as a success rate criterion
for parametrization. A candidate identification error rate was also defined for
adjusting the parameters that are related with the candidate selection stage.
This error rate is set as the number of actual pitches that are not present in the
candidate set divided by the number of actual pitches.
The overall results for the three methods using the random mixtures data
set and the selected parameters are described in Sec. 7.4.2, and the results of
the comparison with other multiple f0 estimation approaches are detailed in
Secs. 7.4.3 and 7.4.4.

Iterative cancellation method

The parameters chosen for the iterative cancellation method are shown in
Tab. 7.1.
In the SLM analysis (see pag. 121), the threshold τ was set to a very
low value, as it is preferable to have a noise peak than discarding a partial
in the preprocessing stage. Different bandwidth values for the SLM were
tested to find the optimal bandwidth W . However, the use of SLM did not
improve the accuracy respect to the systematic selection of all the spectral
peaks (see Fig. 7.11). This can be partially explained because in the test set
there were only harmonic components, making unnecessary to discard spurious
peaks. Besides this reason, the SLM method assumes that there are not two
sinusoidal components closer than W . In some cases, this assumption does not
hold in polyphonic real signals, where typical values of W ∈ [10, 50] Hz exceed
some partial frequency differences. Therefore, the SLM stage was removed to
obtain the results described in Sec. 7.4.2, and the spectral peaks have been

142
 7.4. EVALUATION

Acc
0.4
SLM
Spectral peak picking

0.3

0.2 W
0
10 1
20 2
30 3
40 4
50

Figure 7.11: SLM accuracy respect to the bandwidth W using τ = 0.1, and
comparison with simple spectral peak picking. The other parameters used for
the evaluation are those described in Tab. 7.1.

systematically selected from the magnitude spectrum instead, as SML did not
improve neither the accuracy nor the efficiency with the tested values.
Experimentally, the use of all the spectral peaks yielded exactly the same
results than the selection of those spectral peaks with a magnitude over a low
fixed threshold µ = 0.1. This thresholding, which did not alter the results in the
test set, can reduce the computation time of the overall system to the half21 .
For this reason, this threshold was adopted and subsequently included in the
joint estimation methods.
The overall results without SLM can be seen in Fig. 7.12. In this figure,
the chosen parameter values are in the central intersection, and they correspond
to those described in Tab. 7.1. From these initial values, the parameters have
been changed individually to compare their impact in the accuracy.
The zero padding factor z is useful to accurately identify the frequency of
lower pitches. As shown in Fig. 7.12, the overall accuracy increases importantly
when using zero padding22 (z 6= 20 ). The computational cost derived from the
FFT computation of longer windows must also be taken into account. As the
overall computational cost of this method is very low, a value z = 8, which
slightly improves the accuracy, was chosen.
The range of valid fundamental frequencies comprises the f0 range of the
data set used for the evaluation, therefore it is the same for all the evaluated
methods.
The closest pitch distance value matches the spectral resolution obtained
with zero padding. This way, using a margin of fd = 3 Hz, only spectral
21 Experimentally, the running time was reduced from 146.05 to 73.4 seconds.
22 Due to the FFT constraints, only power of 2 values for z have been tested.

143
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

z
fd
γ
η

z = 23 γ=9 fd = 8
γ=0
η = 0.25

η=0

fd = 1

z = 20

Figure 7.12: Iterative cancellation accuracy adjusting the free parameters.

The abscissae axis is not labeled since these values depend on each particular
parameter (see the first and last values for each parameter to get the grid step).

peaks at ±1 bin from the ideal pitch frequency are considered as f0 candidates.
This parameter increases the accuracy about a 1%. However, as it can be seen
in Fig. 7.12, the value selected for this parameter (fd = 3) is probably too
much restrictive, and a higher range (fd = 5) yields better results. It must be
considered that the iterative cancellation method was developed before having
the random mixtures database, therefore their parameters are not optimally
tuned for this data set. However, experimentally, the accuracy deviation shows
that the chosen values do not differ much from the ones that approximate to
the highest accuracy using this data set.

The postprocessing parameters of the iterative cancellation method are the

minimum note intensity γ and the minimum relative intensity η of a candidate
respect to the maximum intensity of the other simultaneous candidates in the
analyzed frame. The note silence threshold value γ = 5 (equivalent to 18.38 dB)
could not be directly evaluated using the random mixtures data set, as there
are no silences and all the sounds have very similar amplitudes. However, the
results varying γ ∈ [0, 9] show that this value has a low impact in the detection
when there are no silent excerpts in the signal.

144
 7.4. EVALUATION

Stage Parameter Symbol Value

Preprocessing Partial selection threshold µ 0.1
Zero padding factor z 4
Candidate selection Min f0 amplitude ε 2
f0 range [fmin , fmax ] [38, 2100] Hz
Combination generation Max number of candidates F 10
Max polyphony P 6
Partial search bandwidth fr 11 Hz
HPS length H 10
Salience calculation Absolute intensity threshold γ 5
Relative intensity threshold η 0.1
Smoothness weight κ 2
Postprocessing Minimum note duration d 55.68 ms

Table 7.2: Parameters for the joint estimation method I.

The parameter η = 0.1 assumes that the relative intensity of a note is as

minimum a 10% respect to the maximum intensity in the mixture23 . The
effect of this parameter in the detection is shown in Fig. 7.12. It has been
considered that all the notes in the database have similar amplitudes, therefore
this threshold should have a low value for real signals out of the data set, which
are usually of different intensities.

Joint estimation method I

The parameters used for the joint estimation I method are shown in Tab. 7.2,
and the impact in the accuracy when they vary can be seen in Fig. 7.14.
A peak picking threshold µ = 0.1 was chosen, like in the iterative cancellation
method. This parameter increased the efficiency with a very low accuracy cost24 .
Like in the iterative cancellation method, the zero padding factor has shown
to be very relevant to increase the accuracy (see Fig. 7.14). A trade-off value
z = 4, instead of z = 8 used in iterative cancellation method, was chosen to
avoid increasing significantly the computational cost (see Fig. 7.15), which is
higher in this method.
The minimum f0 bin amplitude ε = 2 slightly increases the accuracy and
decreases the candidate selection error (see Fig. 7.13). A higher accuracy was
obtained with ε = 5, but note that this parameter must have a lower value
for the analysis of real musical signals25 , so this more conservative value was
selected instead.
23 It is assumed that, with lower intensities, notes are usually masked by the amplitude of

the other pitches in the mixture, and they can hardly be perceived.
24 Experimentally, in this method the accuracy decreased from 0.553 to 0.548.
25 Unlike in real music, in the test set all the signals had similar and high amplitudes.

145
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

z
ε
fr
F
F =6 z = 20

fr = 8

ε=0

fr = 16
z = 23 ε=5

F = 14

Figure 7.13: Joint estimation method I candidate error rate adjusting the
parameters that have some influence in the candidate selection stage.

z
ε
fr
F
z = 23 H
ε=5 κ
fr = 16

H = 15
F =6 fr = 8 κ=4 F = 14

ε=0

κ=0

H=5

z = 20

Figure 7.14: Joint estimation method I accuracy adjusting the free parameters.

146
 7.4. EVALUATION

z
F
H

F = 14

z = 23 H = 14

H=5

F =6
z = 20

Figure 7.15: Joint estimation method I runtime in seconds adjusting the

parameters that have some influence in the computational cost.

The bandwidth for searching partials fr does not seem to have a great
impact in the accuracy, but it is important in the candidate selection stage
(see Fig. 7.13). An appropriate balance between a high accuracy and a low
candidate selection error rate was obtained using fr = 11 Hz.
The computational cost increases exponentially with the number of candi-
dates F . Therefore, a good choice of F is critical for the efficiency of the method.
Experimentally, F = 10 yielded a good trade-off between the accuracy, the
number of correctly selected candidates and the computational cost (see Figs.
7.13, 7.14 and 7.15).
As previously mentioned, the first partials usually contain most of the energy
of the harmonic sounds. Experimentally, using H = 10 suffices, and higher
values cause low pitches to cancel other higher frequency components. In
addition, note that the computational cost linearly increases with respect to
H.
The smoothness weight which maximizes the accuracy was experimentally
found using κ = 2. It is important to note that without considering spectral
smoothing (κ = 0), the accuracy decreases significantly (see Fig. 7.14).
The postprocessing parameter values for γ and η were selected with the same
values than in the iterative cancellation approach.

147
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

Stage Parameter Symbol Value

Preprocessing Partial selection threshold µ 0.1
Zero padding factor z 4
Candidate selection f0 range [fmin , fmax ] [38, 2100] Hz
Combination generation Max number of candidates F 10
Max polyphony P 6
Partial search bandwidth fr 11 Hz
HPS length H 15
Salience calculation Absolute intensity threshold γ 5
Relative intensity threshold η 0.15
Smoothness weight κ 4
Postprocessing Minimum note duration d 23 ms
(without tracking) Minimum rest duration r 50 ms
Postprocessing Number of adjacent frames K 2
(with tracking)

Table 7.3: Parameters for the joint estimation method II.

An additional experiment was done to measure the importance of the square

factor in Eq. 7.13. Without squaring, the accuracy decreased from 0.548
to 0.476, which is significantly lower, showing the importance of this factor
introduced to favor the sparseness.

Joint estimation method II

The parameters selected for the joint estimation method II are shown in Tab.
7.3. Most of them are the same than in the method I, except from H, η and κ,
which yielded better results with slightly different values (see Fig. 7.16). In the
case of H and κ, the values that maximized the accuracy were selected.
Like in the joint estimation method I, the parameter η = 0.15 has been set
with a conservative value in order to avoid that the system performs worse for
real musical signals, which usually do not have very similar intensities for the
different sounds.
The postprocessing parameters can not be directly evaluated with this
data set, as they have a temporal dimension and each mixture is composed
of a single combination of pitches. However, a value K = 2 (considering 2
previous frames, 2 posterior frames and the target frame) has proven to be
adequate for the analysis of real musical signals out from the data set. This
value provides a notable temporal smoothness without altering significantly the
temporal resolution required for short notes.

148
 7.4. EVALUATION

H
κ
η

η = 0.3
H = 17
κ=6

H = 10 η=0

κ=0

Figure 7.16: Joint estimation method II accuracy respect to the parameters

that changed respect to the method I.

7.4.2 Results using random mixtures

The overall results for the random mixtures data set after the parametrization
stage are described in the following figures of this section (Figs. 7.17 to 7.30).
These results can not be directly compared to the evaluation made by Klapuri
(2006b) using the same data, as in this latter work polyphony estimation and
f0 estimation were evaluated separately (the number of concurrent sounds was
given as a parameter for the pitch estimator), whereas in the present work these
two stages are calculated simultaneously.
As shown in Figs. 7.17 and 7.18, the candidate identification technique used
in the joint estimation method II outperforms the other candidate selection
approaches. It can also be seen (Figs. 7.19 to 7.22) that the joint estimation
method I clearly outperforms the iterative cancellation approach, and the joint
estimation method II gets a higher accuracy than the joint method I.
Respect to the estimation of the number of concurrent sources (Figs. 7.23
to 7.27), the joint estimation method II usually yields better results, but when
there are many simultaneous sources (Fig. 7.26), it tends to underestimate
the number of concurrent sounds, probably due to the combination of adjacent
frames. Looking at the evaluation in function of the pitch (Figs. 7.28 to 7.30),
it can be seen that the best results are located in the central pitch range.

149
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

1
Iterative
Joint I
Joint II
0.8

0.6

0.4

0.2

0
01 12 24 36 4

Figure 7.17: Candidate identification error rate with respect to the polyphony
(1, 2, 4 and 6 simultaneous pitches) of the ground truth mixtures.

1
Iterative
Joint I
Joint II
0.8

0.6

0.4

0.2

0
0

Figure 7.18: Global candidate identification error rate.

150
 7.4. EVALUATION

1
Precision
Recall
Accuracy
0.8

0.6

0.4

0.2

0
01 12 24 36

Figure 7.19: Pitch detection results for the iterative cancellation method with
respect to the ground-truth mixtures polyphony.

1
Precision
Recall
Accuracy
0.8

0.6

0.4

0.2

0
01 12 24 36

Figure 7.20: Pitch detection results for the joint estimation method I with
respect to the ground-truth mixtures polyphony.

151
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

1
Precision
Recall
Accuracy
0.8

0.6

0.4

0.2

0
01 12 24 36

Figure 7.21: Pitch detection results for the joint estimation method II with
respect to the ground-truth mixtures polyphony.

1
Precision
Recall
Accuracy
0.8

0.6

0.4

0.2

0
0
Iterative cancellation Joint1method I Joint 2method II 3

Figure 7.22: Comparison of the global pitch detection results for the three
methods.

152
 7.4. EVALUATION

1000
Iterative
Joint I
Joint II
800

600

400

200

0
0 1 2 3 4 5 6 7 8 9

Figure 7.23: Number of concurrent sounds estimated for a single source.

1000
Iterative
Joint I
Joint II
800

600

400

200

0
0 1 2 3 4 5 6 7 8 9

Figure 7.24: Number of concurrent sounds estimated for two simultaneous

sources.

153
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

1000
Iterative
Joint I
Joint II
800

600

400

200

0
0 1 2 3 4 5 6 7 8 9

Figure 7.25: Number of concurrent sounds estimated for four simultaneous

sources.

1000
Iterative
Joint I
Joint II
800

600

400

200

0
0 1 2 3 4 5 6 7 8 9

Figure 7.26: Estimation of the number of concurrent sources for six

simultaneous sources.

154
 7.4. EVALUATION

1
Iterative
Joint I
Joint II
0.8

0.6

0.4

0.2

0
0

Figure 7.27: Global polyphony estimation results.

Figure 7.28: Precision, recall and accuracy of the iterative cancellation method
in function of the MIDI pitch number.

155
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

Figure 7.29: Precision, recall and accuracy of the joint estimation method I in
function of the MIDI pitch number.

Figure 7.30: Precision, recall and accuracy of the joint estimation method II
in function of the MIDI pitch number.

156
 7.4. EVALUATION

7.4.3 Evaluation and comparison with other methods

In order to evaluate the proposed methods using real musical signals and to
compare them with other approaches, the iterative cancellation algorithm and
the joint estimation method I were submitted to the MIREX (2007) multiple
f0 estimation and tracking contest, whereas the joint estimation method II was
evaluated in MIREX (2008).
The data set used in MIREX (2007) and MIREX (2008) were essentially
the same, consisting of a woodwind quintet transcription of the fifth variation
from Beethoven, plus some synthesized pieces using Goto (2003) samples, and
polyphonic piano recordings using a diskclavier piano. There were clips of
30 seconds for each polyphony (2-3-4-5), for a total of 30 examples, plus 10
polyphonic piano pieces of 30 seconds. The details of the ground-truth labelling
are described in (Bay et al., 2009).
The MIREX evaluation was done at two different levels; frame by frame pitch
estimation and note tracking. The first mode evaluates the correct detection in
isolated frames, whereas the second task also considers the temporal coherence
of the detection.
For the frame level task, evaluation of the active pitches is done every 10 ms.
For this reason, the hop size of the joint estimation methods26 were set to
obtain an adequate temporal resolution. Precision, recall, and accuracy were
reported. A returned pitch is assumed to be correct if it is within a half semitone
of a ground-truth pitch for that frame. Only one ground-truth pitch can be
associated with each returned pitch. The error metrics from Poliner and Ellis
(2007a) and previously described in pag. 51 were also used in the evaluation.
For the note tracking task, precision, recall, and F-measure were reported. A
ground-truth note is assumed to be correctly transcribed if the method returns
a note that is within a half semitone of that note, the yielded note onset is
within a ±50 ms range of the onset of the ground truth note, and its offset is
within 20% range of the ground truth note offset. One ground truth note can
only be associated with one transcribed note.
The data set is not publicly available, therefore the experiments using these
data can not be replicated out of the MIREX contests.

Iterative cancellation method

The iterative cancellation approach does not perform a frame by frame

evaluation, as it uses only those frames that are after the detected onsets to
yield the pitches for each inter-onset interval. Although it does not perform f0
tracking, onset times provide the indications about the beginning of the notes,
therefore it was only submitted to the note tracking task.
26 The iterative cancellation method was only presented to the note tracking task.

157
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

Participant Runtime (sec) Machine

Iterative cancellation 165 ALE Nodes
Joint estimation method I 364 ALE Nodes
AC3 900 MAC
AC4 900 MAC
EV4 2475 ALE Nodes
EV3 2535 ALE Nodes
KE3 4140 ALE Nodes
PE2 4890 ALE Nodes
RK 3285 SANDBOX
KE4 20700 ALE Nodes
VE 390600 ALE Nodes

Table 7.4: MIREX (2007) note tracking runtimes. Participant, running time
(in seconds), and machine where the evaluation was performed are shown.

The parameters of the iterative cancellation method are those described in

Tab. 7.1, using SLM in the preprocessing stage. Although the joint estimation
method I is a pure frame by frame estimation approach, it was also submitted
for the note tracking task, to compare the results with the iterative approach
at least at note level.
The overall results using onset and pitch27 are shown in Tab. 7.5, and the
runtimes are in Tab. 7.4. The iterative cancellation method was the one with the
lowest computational cost. As expected, the F-measure was not very high, but
the method outperformed two complex approaches. However, joint estimation
method I, which is really a frame by frame estimation method, obtained better
results than the iterative cancellation approach for the note tracking task.

Joint estimation method I

The joint estimation method I was submitted for evaluation in MIREX (2007)
frame by frame and note tracking contests with the parameters specified in
Tab. 7.2.
The results for the frame by frame analysis are shown in Tab. 7.6, and the
corresponding runtimes in Tab. 7.7. The accuracy of this method was close28
to the highest accuracy among the evaluated methods, being the one with the
highest precision and the lowest Etot error. The precision, recall and accuracy
were slightly better than those obtained with the random mixtures database.
Probably, this is because the random mixtures database spans a pitch range
wider than the MIREX data set, and very low or very high frequency pitches
are those harder to detect (see Fig. 7.29).
27 Evaluation for onset, offset and pitch were also done in MIREX (2007), but results are

not reported in this work, as the iterative estimation system does not consider offsets.
28 About 0.025 lower.

158
 id Participant Method Avg. F-m Prec Rec Avg. Overlap
RK Ryynänen and Klapuri (2005) Iterative cancellation + HMM tracking 0.614 0.578 0.678 0.699
EV4 Vincent et al. (2007) Unsupervised learning (NMF) 0.527 0.447 0.692 0.636
PE2 Poliner and Ellis (2007a) Supervised learning (SVM) 0.485 0.533 0.485 0.740
EV3 Vincent et al. (2007) Unsupervised learning (NMF) 0.453 0.412 0.554 0.622
PI2 Pertusa and Iñesta (2008a) Joint estimation method I 0.408 0.371 0.474 0.665
KE4 Kameoka et al. (2007) Statistical spectral models 0.268 0.263 0.301 0.557
KE3 Kameoka et al. (2007) Statistical spectral models 0.246 0.216 0.323 0.610
PI3 Lidy et al. (2007) Iterative cancellation 0.219 0.203 0.296 0.628
VE2 Emiya et al. (2007, 2008b) Joint estimation + Bayesian models 0.202 0.338 0.171 0.486
AC4 Cont (2007) Unsupervised learning (NMF) 0.093 0.070 0.172 0.536
AC3 Cont (2007) Unsupervised learning (NMF) 0.087 0.067 0.137 0.523

Table 7.5: MIREX (2007) note tracking results based on onset and pitch. Average F-measure, precision, recall, and average
overlap are shown for each participant.

159
7.4. EVALUATION
160
id Participant Method Acc Prec Rec Etot Esubs Emiss Ef a
RK Ryynänen and Klapuri (2005) Iterative cancellation + HMM tracking 0.605 0.690 0.709 0.474 0.158 0.133 0.183
CY Yeh (2008) Joint estimation 0.589 0.765 0.655 0.460 0.108 0.238 0.115
ZR Zhou et al. (2009) Salience function (RTFI) 0.582 0.710 0.661 0.498 0.141 0.197 0.160
PI1 Pertusa and Iñesta (2008a) Joint estimation method I 0.580 0.827 0.608 0.445 0.094 0.298 0.053
EV2 Vincent et al. (2007) Unsupervised learning (NMF) 0.543 0.687 0.625 0.538 0.135 0.240 0.163
CC1 Cao et al. (2007) Iterative cancellation 0.510 0.567 0.671 0.685 0.200 0.128 0.356
SR Raczynski et al. (2007) Unsupervised learning (NNMA) 0.484 0.614 0.595 0.670 0.185 0.219 0.265
EV1 Vincent et al. (2007) Unsupervised learning (NMF) 0.466 0.659 0.513 0.594 0.171 0.371 0.107
PE1 Poliner and Ellis (2007a) Supervised learning (SVM) 0.444 0.734 0.505 0.639 0.120 0.375 0.144
PL Leveau (2007) Matching pursuit 0.394 0.689 0.417 0.639 0.151 0.432 0.055
CC2 Cao et al. (2007) Iterative cancellation 0.359 0.359 0.767 1.678 0.232 0.001 1.445
KE2 Kameoka et al. (2007) Statistical spectral models (HTC) 0.336 0.348 0.546 1.188 0.401 0.052 0.734
KE1 Kameoka et al. (2007) Statistical spectral models (HTC) 0.327 0.335 0.618 1.427 0.339 0.046 1.042
AC2 Cont (2007) Unsupervised learning (NMF) 0.311 0.373 0.431 0.990 0.348 0.221 0.421
AC1 Cont (2007) Unsupervised learning (NMF) 0.277 0.298 0.530 1.444 0.332 0.138 0.974
VE Emiya et al. (2007, 2008b) Joint estimation + Bayesian models 0.145 0.530 0.157 0.957 0.070 0.767 0.120

Table 7.6: MIREX (2007) frame by frame evaluation results. Accuracy, precision, recall, and the error metrics proposed by
Poliner and Ellis (2007a) are shown for each participant.
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS
 7.4. EVALUATION

id Runtime (sec) Machine

ZR 271 BLACK
Joint estimation method I 364 ALE Nodes
AC1 840 MAC
AC2 840 MAC
EV2 2233 ALE Nodes
EV1 2366 ALE Nodes
CC1 2513 ALE Nodes
CC2 2520 ALE Nodes
RK 3540 SANDBOX
PE1 4564 ALE Nodes
PL 14700 ALE Nodes
KE2 19320 ALE Nodes
KE1 38640 ALE Nodes
SR 41160 ALE Nodes
CY 132300 ALE Nodes
VE 364560 ALE Nodes

Table 7.7: MIREX (2007) frame by frame runtimes. The first column shows
the participant, the second is the runtime and the third column is the machine
where the evaluation was performed. ALE Nodes was the fastest machine.

The method was also evaluated in the note tracking contest. Despite it was
not designed for this task, as the analysis is performed without information of
neighboring frames but converting consecutive pitch detections into notes, the
results were not bad, as shown in Tab. 7.5.
The joint estimation method I was also very efficient respect to the other
state of the art methods presented (see Tab. 7.7), specially considering that it
is a joint estimation approach.

Joint estimation method II

The joint estimation method II was submitted to MIREX (2008) for frame by
frame and note tracking evaluation. The method was presented for both tasks
in two setups: with and without f0 tracking.
The difference between using f0 tracking or not is the postprocessing stage
(see Tab. 7.3). In the first setup, notes shorter than a minimum duration are
just removed, and when there are short rests between two consecutive notes
of the same pitch, the notes are merged. Using f0 tracking, the methodology
described in Sec. 7.3.3 is performed instead, increasing the temporal coherence
of the estimate with the wDAG.
Experimentally, the joint estimation method II was very efficient compared
to the other approaches presented, as shown in Tabs. 7.8 and 7.9.
The results for the frame by frame task can be seen in Tab. 7.10. The
accuracy for the joint estimation method II without f0 tracking is satisfactory,

161
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

Participant Runtime (sec)

MG 99
Joint estimation II 792
Joint estimation II + tracking 955
VBB 2081
CL1 2430
CL2 2475
RK 5058
EOS 9328
DRD 14502
EBD1 18180
EBD2 22270
YRC1 57483
YRC2 57483
RFF2 70041
RFF1 73784

Table 7.8: MIREX (2008) frame by frame runtimes. Participants and runtimes
are shown. All the methods except MG were evaluated using the same machine.

Participant Runtime (sec)

Joint estimation II 790
ZR3 871
Joint estimation II + tracking 950
ZR1 1415
ZR2 1415
VBB 2058
RK 5044
EOS 9328
EBD1 18180
EBD2 22270
YRC 57483
RFF2 71360
RFF1 73718

Table 7.9: MIREX (2008) note tracking runtimes. Participants and runtimes
are shown. All the methods except ZR were evaluated using the same machine.

and the method obtained the highest precision and the lowest Etot error among
all the analyzed approaches.
The inclusion of f0 tracking did not improve the results for frame by
frame estimation, but in the note tracking task (see Tab. 7.11), the results
outperformed those obtained without tracking.

7.4.4 Overall MIREX comparison

As the ground-truth used for MIREX (2007) and MIREX (2008) multiple f0
estimation and tracking contest were the same. In the review from Bay et al.
(2009), the results of the algorithms evaluated in both MIREX editions are
analyzed.

162
 id Participant Method Acc Prec Rec Etot Esubs Emiss Ef a
YRC2 Yeh et al. (2008) Joint estimation + f0 tracking 0.665 0.741 0.780 0.426 0.108 0.127 0.190
YRC1 Yeh et al. (2008) Joint estimation 0.619 0.698 0.741 0.477 0.129 0.129 0.218
PI2 Pertusa and Iñesta (2008b) Joint estimation II 0.618 0.832 0.647 0.406 0.096 0.257 0.053
RK Ryynänen and Klapuri (2005) Iterative cancellation + HMM tracking 0.613 0.698 0.719 0.464 0.151 0.130 0.183
PI1 Pertusa and Iñesta (2008b) Joint estimation II + tracking 0.596 0.824 0.625 0.429 0.101 0.275 0.053
VBB Vincent et al. (2007) Unsupervised learning (NMF) 0.540 0.714 0.615 0.544 0.118 0.267 0.159
DRD Durrieu et al. (2008) Iterative cancellation 0.495 0.541 0.660 0.731 0.245 0.096 0.391
CL2 Cao and Li (2008) Iterative cancellation 0.487 0.671 0.560 0.598 0.148 0.292 0.158
EOS Egashira et al. (2008) Statistical spectral models (HTC) 0.467 0.591 0.546 0.649 0.210 0.244 0.194
EBD2 Emiya et al. (2008a) Joint estimation + Bayesian models 0.452 0.713 0.493 0.599 0.146 0.362 0.092
EBD1 Emiya et al. (2008a) Joint estimation + Bayesian models 0.447 0.674 0.498 0.629 0.161 0.341 0.127
MG Groble (2008) Database matching 0.427 0.481 0.570 0.816 0.298 0.133 0.385
CL1 Cao and Li (2008) Iterative cancellation 0.358 0.358 0.763 1.680 0.236 0.001 1.443
RFF1 Reis et al. (2008a) Supervised learning (genetic) 0.211 0.506 0.226 0.854 0.183 0.601 0.071
RFF2 Reis et al. (2008a) Supervised learning (genetic) 0.183 0.509 0.191 0.857 0.155 0.656 0.047

Table 7.10: MIREX (2008) frame by frame evaluation results. Accuracy, precision, recall, and the error metrics proposed by
Poliner and Ellis (2007a) are shown for each method.

163
7.4. EVALUATION
164
id Participant Method Avg. F-m Prec Rec Avg. Overlap
YRC Yeh et al. (2008) Joint estimation + f0 tracking 0.355 0.307 0.442 0.890
RK Ryynänen and Klapuri (2005) Iterative cancellation + HMM tracking 0.337 0.312 0.382 0.884
ZR3 Zhou and Reiss (2008) Salience function (RTFI) 0.278 0.256 0.314 0.874
ZR2 Zhou and Reiss (2008) Salience function (RTFI) 0.263 0.236 0.306 0.874
ZR1 Zhou and Reiss (2008) Salience function (RTFI) 0.261 0.233 0.303 0.875
PI1 Pertusa and Iñesta (2008b) Joint estimation II + tracking 0.247 0.201 0.333 0.862
EOS Egashira et al. (2008) Statistical spectral models (HTC) 0.236 0.228 0.255 0.856
VBB Vincent et al. (2007) Unsupervised learning (NMF) 0.197 0.162 0.268 0.829
PI2 Pertusa and Iñesta (2008b) Joint estimation II 0.192 0.145 0.301 0.854
EBD1 Emiya et al. (2008a) Joint estimation + Bayesian models 0.176 0.165 0.200 0.865
EBD2 Emiya et al. (2008a) Joint estimation + Bayesian models 0.158 0.153 0.178 0.845
RFF2 Reis et al. (2008a) Supervised learning (genetic) 0.032 0.037 0.030 0.645
RFF1 Reis et al. (2008a) Supervised learning (genetic) 0.028 0.034 0.025 0.683

Table 7.11: MIREX (2008) note tracking results based on onset, offset, and pitch. Average F-measure, precision, recall, and
average overlap are shown for each method.
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS
 7.4. EVALUATION

Figure 7.31: Fig. from Bay et al. (2009), showing Esubs , Emiss and Ef a for
all MIREX 2007 and MIREX 2008 multiple fundamental frequency estimation
methods ordered by Etot . PI2-08 is the joint estimation method II without
tracking, PI1-08 is the same method with tracking, and PI-07 is the joint
estimation method I.

As shown in Fig. 7.32, the proposed joint estimation methods achieve a

high overall accuracy and the highest precision rates among all the reference
methods. The joint estimation method II also obtained the lowest error (Etot )
rate from the 31 methods submitted in both editions (see Fig. 7.31) using the
metric proposed by Poliner and Ellis (2007a).
In the evaluation of note tracking considering only onsets, the proposed
methods showed lower accuracies (Fig. 7.33), as only the joint estimation
method II can perform a very basic f0 tracking. With respect to the
iterative cancellation approach, the accuracy was lower than the joint estimation
methods, but not disappointing when compared to the other algorithms
evaluated, given that this is a very simple method mainly intended for piano
music.
As pointed out by Bay et al. (2009), the reason behind the different results in
Etot and accuracy in the frame by frame evaluation (see Figs. 7.31 and 7.32) is

165
166
Figure 7.32: Fig. from Bay et al. (2009). Precision, recall and overall accuracy for all MIREX 2007 and MIREX 2008 multiple
fundamental frequency estimation methods ordered by accuracy. PI2-08 is the joint estimation method II without tracking,
PI1-08 is the same method with tracking, and PI-07 is the joint estimation method I.
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS
 7.4. EVALUATION

Figure 7.33: Fig. from Bay et al. (2009). Precision, recall, average F-measure
and average overlap based on note onset for MIREX 2007 and MIREX 2008 note
tracking subtask. PI2-08 is the joint estimation method II without tracking,
PI1-08 is the same method with tracking, PI1-07 is the joint estimation method
I and PI2-07 is the iterative cancellation method.

that most of the reported f0 were correct, but multiple f0 estimation algorithms
tend to under-report and miss many active f0 in the ground-truth.
While the proposed joint estimation methods I and II achieved the lowest
Etot score, there are very few false alarms compared to miss errors. On the other
hand, the methods from Ryynänen and Klapuri (2005) and Yeh et al. (2008)
have a better balanced precision, recall, as well as a good balance in the three
error types, and as a result, have the highest accuracies for MIREX (2007) and
MIREX (2008), respectively.
Citing Bay et al. (2009), ”Inspecting the methods used and their perfor-
mances, we can not make generalized claims as to what type of approach works
best. In fact, statistical significance testing showed that the top three methods29
were not significantly different.”

29 (Yeh et al., 2008; Pertusa and Iñesta, 2008b; Ryynänen and Klapuri, 2005).

167
7. MULTIPLE F0 ESTIMATION USING SIGNAL PROCESSING METHODS

7.5 Conclusions
In this chapter, three different signal processing methods have been proposed for
multiple f0 estimation. Unlike the supervised learning approaches previously
described, these signal processing schemes can be used to transcribe real music
without any a-priori knowledge of the sources.
The first method is based on iterative cancellation, and it is a simple
approach which is mainly intended for the transcription of piano sounds at a low
computational cost. For this reason, only one frame in an inter-onset interval
is analyzed, and the interaction between harmonic sources is not considered.
A fixed spectral pattern is used to subtract the harmonic components of each
candidate.
The joint estimation method I introduces a more complex methodology.
The spectral patterns are inferred from the analysis of different hypotheses
taking into account the interactions with the other sounds. The combination
of harmonic patterns that maximizes a criterion based on the sum of harmonic
amplitudes and spectral envelope smoothness is chosen at each frame.
The third method extends the previous joint estimation method considering
adjacent frames for adding temporal smoothing. This method can be com-
plemented with a f0 tracking stage, using a weighted direct acyclic graph, to
increase the temporal coherence of the detection.
The proposed methods have been evaluated and compared to other works.
The iterative cancellation approach, mainly intended for piano transcription, is
very efficient and it has been successfully used for genre classification and other
MIR tasks (Lidy et al., 2007) with computational cost restrictions.
The joint estimation methods obtained a high accuracy and the lowest Etot
among all the multiple f0 algorithms submitted in MIREX (2007) and MIREX
(2008). Although all possible combinations of candidates are evaluated at each
frame, the proposed approaches have a very low computational cost, showing
that it is possible to make an efficient joint estimation method.
Probably, the f0 tracking stage added to the joint estimation method II is
too simple, and it should be replaced by a more reliable method in a future
work. For instance, the transition weights could be learned from a labeled test
set, or a more complex f0 tracking method like the high-order HMM scheme
from Chang et al. (2008) could be used instead. Besides intensity, the centroid
of an HPS should also have a temporal coherence when belonging to the same
source, therefore this parameter could also be considered for tracking.
Using stochastic models, a probability can be assigned to each pitch in order
to remove those that are less probable given their context. For example, in
a melodic line it is very unlikely that a non-diatonic note two octaves higher
or lower than its neighbours appears. Musical probabilities can be taken into

168
 7.5. CONCLUSIONS

account, like in (Ryynänen and Klapuri, 2005), to remove very unprobable notes.
The adaptation to polyphonic music of the stochastic approach from Pérez-
Sancho (2009) is also planned as future work, in order to use it in the multiple
f0 estimation methods to obtain a musically coherent detection.
The evaluation and further research of the alternative architectures proposed
for the joint estimation method II (see Sec. 7.3.4) is also left for future work.

169
 Conclusions and future work
8
This work has addressed the automatic music transcription problem using
different strategies. Efficient novel methods have been proposed for onset
detection and multiple f0 estimation, using supervised learning and signal
processing techniques. The main contributions of this work can be summarized
in the following points:

• An extensive review of the state of the art methods for onset detection
and multiple f0 estimation. The latter methods have been classified
into salience functions, iterative cancellation, joint estimation, supervised
learning, unsupervised learning, matching pursuit, Bayesian models,
statistical spectral models, blackboard systems, and database matching
methods. An analysis of the strengths and limitations for each category
has also been done.

• The development of an efficient approach for onset detection and the

construction of a ground-truth data set for this task. The main novelties in
this field are the use of a 1/12 octave filter bank to compress the harmonic
information and the simple onset detection functions proposed. The
presented method is mainly intended for percussive onset detection, as it
detects energy abrupt energy changes, but it also considers the properties
of harmonic sounds, making it robust against spectral variations produced
during the sustain stage of the sounds. The algorithm was evaluated and
compared to other works yielding promising results.

• Two novel approaches for multiple pitch estimation of a priori known

sounds using supervised learning methods. These algorithms were one
of the first machine learning methods proposed for this task. A harmonic
filter bank was used to reduce the amount of spectral information to feed a
time-delay neural network (TDNN), while preserving the main harmonic
content. A ground-truth data set of synthetic sounds was generated to

171
8. CONCLUSIONS AND FUTURE WORK

evaluate the method. The conclusions extracted from the comparison

between the k nearest neighbors (kNN) and the time-delay neural networks
for this task are also interesting. The TDNN clearly outperformed the
results obtained by the kNN using synthesized mixtures, showing the
advantages of the network for generalization within a large observable
space. Alternative activation functions to generalize the kNN prototypes
were also proposed, but the results were still far from those obtained with
the TDNN.

• A simple iterative cancellation approach, mainly intended to transcribe

piano music at a low computational cost. A complete system to load
an audio file and generate a MIDI file was developed. A novel scheme,
based on the analysis of the spectra only after each onset to improve
the efficiency, was proposed, and a sinusoidal likeness measure was also
investigated for this task. This method was the basis for the subsequent
joint estimation approaches, and it has been successfully used for genre
classification and other MIR tasks.

• Heuristic multiple f0 algorithms based on signal processing to analyze real

music without any a priori knowledge. These methods, which are probably
the main contribution of this thesis, experimentally reached the state of the
art for this task with a high efficiency. The harmonic patterns are inferred
from the spectrum considering intensity and a novel smoothness measure
in a joint scheme which takes into account the source interactions. The
proposed approach also introduced a novel temporal smoothing technique
by considering the pitch combinations in adjacent frames. A simple f0
tracking method is also introduced using a weighted direct acyclic graph.
These novel methods achieved high success rates, but also with a very high
efficiency, which is the main handicap of joint estimation approaches.

8.1 Discussion and future lines of work

The proposed TDNN scheme could easily be adapted for onset detection in a
future work. Like in the machine learning methods described for multiple pitch
estimation, the 1/12 octave filter bank can be used to obtain the network input
data, but for this application only an output neuron should be necessary to
classify each frame as onset or not-onset. This learning scheme also implies to
extend the compiled onset detection data set in order to get a larger and reliable
training set.
The supervised learning method should also be trained and evaluated using
real mixtures of different timbres and real audio signals aligned with the ground-
truth pitches. As previously discussed, it is not an easy task to get an aligned

172
 8.1. DISCUSSION AND FUTURE LINES OF WORK

database. However, a similar scheme that proposed by (Yeh (2008)) could be

used to build the training set. In the latter work, MIDI files are splitted into
several files containing tracks of separate notes. Then, they are individually
synthesized and the f0 are estimated using the YIN algorithm from de Cheveigné
and Kawahara (2002). The f0 of the individual note samples collectively
establish the ground truth of the synthesized polyphonic signal.
Probably, multiple f0 estimation methods that only consider individual
frames can hardly outperform the current approaches. The lack of data within
that short period is not enough to detect the pitches even for an expert musician.
Context plays an important role in music. For instance, it is very hard to detect
the pitches when listening two songs at different tempo playing simultaneously,
even when they are not very complex.
Therefore, pitch estimation should be complemented with temporal infor-
mation in some way. The coherence of the detections along time has been
considered in the joint estimation method II, but it could be extended using a
reliable f0 tracking method. However, as pointed out by Yeh (2008), f0 tracking
in a joint manner is complicated. A robust f0 tracking algorithm should analyze
many possible simultaneous pitch combinations at each frame for a long term,
and this is a challenging task from a computational point of view. The problem
can be simplified by tracking individual candidate trajectories like in the high-
order HMM method from Chang et al. (2008).
The further research of the alternative schemes proposed for the joint
estimation method II is also a promising research subject. The methodology
of this approach allows, for instance, to merge combinations of those frames
that are between two consecutive onsets, yielding the pitches within the inter-
onset interval. Perceptually, the results obtained with this scheme were better
than in the analysis of a few adjacent frames. However, the lower temporal
resolution and the errors of the onset detection method, and the problem with
the offsets in the inter-onset interval, condition the success rate using a classical
frame by frame evaluation metric. As future work, it is planned to evaluate
these schemes using a perceptual metric.
Multimodal information is also important for future research subjects. The
inclusion of musical models by considering tonality, tempo, or meter to infer
pitch probabilities could complement the pitch estimates. These lines of work are
planned within the DRIMS project in collaboration with the Music Technology
Group from the Universitat Pompeu Fabra.
Interactive music transcription is also planned as future work within the
MIPRCV (Consolider Ingenio 2010) project. The main goal is to develop a
computer-assisted method for music transcription, similarly to the system from
Vidal et al. (2006) for machine translation. Using a visual interface, the portions
of automatically transcribed music can be accepted or amended by an expert

173
8. CONCLUSIONS AND FUTURE WORK

musician. Then, these user-validated portions can be used by the multiple

f0 estimation method to produce further, hopefully improved suggestions by
correcting frequent errors.
Further research of the proposed methods applied to other MIR tasks like
genre or mood classification is also currently investigated in collaboration with
the Department of Software Technology and Interactive Systems from the
Vienna University of Technology.

8.2 Publications
Some contents of this thesis have been published in journals and conference
proceedings. Here is a list of publications in chronological order.

• Pertusa, A. and Iñesta, J. M. (2004). Pattern recognition algorithms for

polyphonic music transcription. In Fred, A., editor, Pattern Recognition in
Information Systems (PRIS), pages 80-89, Porto, Portugal. [Chapter 6]

• Pertusa, A., Klapuri, A., and Iñesta, J. M. (2005). Recognition of note

onsets in digital music using semitone bands. Lecture Notes in Computer
Science, 3773:869-879. [Chapter 5]

• Pertusa, A. and Iñesta, J. M. (2005). Polyphonic monotimbral music

transcription using dynamic networks. Pattern Recognition Letters,
26(12):1809-1818. [Chapter 6]

• Lidy, T., Rauber, A., Pertusa, A., and Iñesta, J. M. (2007). Improving
genre classification by combination of audio and symbolic descriptors using
a transcription system. In Proc. of the 8th International Conference
on Music Information Retrieval (ISMIR), pages 61-66, Vienna, Austria.
[Chapter 7]

• Pertusa, A. and Iñesta, J. M. (2007). Multiple fundamental frequency

estimation based on spectral pattern loudness and smoothness. In MIREX
(2007), multiple f0 estimation and tracking contest. [Chapter 7]

• Lidy, T., Rauber, A., Pertusa, A., Ponce de León, P. J., and Iñesta, J.
M. (2008). Audio music classification using a combination of spectral,
timbral, rhythmic, temporal and symbolic features. In MIREX (2008),
audio genre classification contest, Philadelphia, PA. [Chapter 7]

• Pertusa, A. and Iñesta, J. M. (2008). Multiple fundamental frequency

estimation using Gaussian smoothness and short context. In MIREX
(2008), multiple f0 estimation and tracking contest. [Chapter 7]

174
 8.2. PUBLICATIONS

• Pertusa, A. and Iñesta, J. M. (2008). Multiple fundamental frequency

estimation using Gaussian smoothness. In Proc. of the IEEE Int. Conf.
on Acoustics, Speech, and Signal Processing (ICASSP), pages 105-108,
Las Vegas, NV. [Chapter 7]

• Lidy, T., Grecu, A., Rauber, A., Pertusa, A., Ponce de León, P. J., and
Iñesta, J. M. (2009). A multi-feature multi-classifier ensemble approach
for audio music classification. In MIREX (2009), audio genre classification
contest, Kobe, Japan. [Chapter 7]

• Pertusa, A. and Iñesta, J. M. (2009). Note onset detection using one

semitone filter-bank for MIREX 2009. In MIREX (2009), onset detection
contest, Kobe, Japan. [Chapter 5]

175
Summary in Spanish required by

A
the regulations of the University of
Alicante.

Resumen

Agradecimientos
Antes de nada, me gustarı́a agradecer a todos los miembros del grupo de música
por ordenador de la Universidad de Alicante por proporcionar una excelente
atmósfera de trabajo. Especialmente, al coordinador del grupo y supervisor de
este trabajo, José Manuel Iñesta. Su incansable espı́ritu cientı́fico proporciona
un marco de trabajo excelente para inspirar las nuevas ideas que nos hacen
crecer y avanzar continuamente. Este trabajo no hubiera sido posible sin su
consejo y ayuda.
Escribir una tesis no es una tarea fácil sin la ayuda de mucha gente.
Primero, me gustarı́a agradecer a toda la plantilla de nuestro Grupo de
Reconocimiento de Formas e Inteligencia Artificial (GRFIA) y, en general,
a todo el Departamento de Lenguajes y Sistemas Informáticos (DLSI) de la
Universidad de Alicante. Mis estancias de investigación con el Audio Research
Group, Tampere University of Technology (Tampere), Music Technology Group
(MTG), Universitat Pompeu Fabra (Barcelona) y Department of Software
Technology and Interactive Systems, Vienna University of Technology (Viena),
también han contribuido notablemente a la realización de este trabajo. He
crecido mucho, como cientı́fico y como persona, aprendiendo de los integrantes
de estos centros de investigación.
También me gustarı́a agradecer a la gente que ha contribuido directamente a
este trabajo. A Francisco Moreno, por retrasar algunas de mis responsabilidades
docentes durante la escritura de este documento y por proporcionar el código
de los algoritmos de k vecinos más cercanos. He aprendido la mayorı́a de
las técnicas que conozco para transcripción musical de Anssi Klapuri. Estaré
eternamente agradecido por los grandes momentos que pasé en Tampere y por su
generosa acogida. Anssi ha contribuido directamente a esta tesis proporcionando
las bases para el código de similitud sinusoidal y la base de datos de acordes
aleatorios que han posibilitado la evaluación y la mejora de los algoritmos

177
A. RESUMEN

propuestos. En este trabajo también ha colaborado directamente uno de mis

estudiantes de grado, Jasón Box, que ha construido la base de datos ODB y ha
traducido de C++ a D2K el código de detección de inicios de eventos musicales.
También me gustarı́a expresar mi gratitud a los revisores y a todos los
miembros del tribunal.
Este trabajo no habrı́a sido posible sin la financiación del proyecto PROS-
EMUS1 y el programa de investigación Consolider Ingenio 2010 MIPRCV2 .
También ha sido financiado por los proyectos españoles CICYT TAR3 y TIRIG4 ,
y por los fondos FEDER de la Unión Europea y los proyectos de la Generalitat
Valenciana GV04B-541 y GV06/166.
Dejando a un lado la investigación, me gustarı́a agradecer a mi familia y
amigos (no voy a nombraros porque ya sabéis quiénes sois). A pesar de que no
saben exactamente en qué trabajo y nunca se leerán un documento cientı́fico
aburrido (y en inglés), su amistad y generosidad han sido cruciales durante este
largo periodo.
Por último, esta tesis está dedicada a la persona más importante de mi vida,
Teima, por su cariño y paciencia durante todo este tiempo.

1- Introducción6
La transcripción musical automática consiste en extraer las notas que están
sonando (la partitura) a partir de una señal de audio digital. En el caso de la
transcripción polifónica, se parte de señales de audio que pueden contener varias
notas sonando simultáneamente.
Una partitura es una guı́a para interpretar información musical, y por tanto
puede representarse de distintas maneras. La representación más extendida
es la notación moderna usada en música tonal occidental. Para extraer una
representación comprensible en dicha notación, además de las notas, sus tiempos
de inicio y sus duraciones, es necesario indicar el tempo, la tonalidad y la
métrica.
La aplicación más obvia de la extracción de la partitura es ayudar a un
músico a escribir la notación musical a partir del sonido, lo cual es una tarea
complicada cuando se hace a mano. Además de esta aplicación, la transcripción
automática también es útil para otras tareas de recuperación de información
musical, como detección de plagios, identificación de autor, clasificación de
género, y asistencia a la composición cambiando la instrumentación o las notas
para generar nuevas piezas musicales a partir de una ya existente. En general,
1 Código TIN2006-14932-C02
2 Código CSD2007-00018
3 Código TIC2000-1703-CO3-02
4 Código TIC2003-08496-C04
6 Introduction.

178
estos algoritmos también pueden proporcionar información sobre las notas para
aplicar métodos que trabajan sobre música simbólica.
La transcripción musical automática es una tarea de recuperación de
información musical en la que están implicadas varias disciplinas, tales como
el procesamiento de señales, el aprendizaje automático, la informática, la
psicoacústica, la percepción musical y la teorı́a musical.
Esta diversidad de factores provoca que haya muchas formas de abordar
el problema. La mayorı́a de trabajos previos han utilizado diversos enfoques
dentro del campo del procesamiento de la señal, aplicando metodologı́as para
el análisis en el dominio de la frecuencia. En la literatura podemos encontrar
múltiples algoritmos de separación de señales, sistemas que emplean algoritmos
de aprendizaje y clasificación para detectar las notas, enfoques que consideran
modelos psicoacústicos de la percepción del sonido, o sistemas que aplican
modelos musicológicos como medida de coherencia de la detección.
La parte principal de un sistema de transcripción musical es el sistema
de detección de frecuencias fundamentales, que determina el número de notas
que están sonando en cada instante, sus alturas y sus tiempos de activación.
Además del sistema de detección de frecuencias fundamentales, para obtener
la transcripción completa de una pieza musical es necesario estimar el tempo
a través de la detección de pulsos musicales, y obtener el tipo de compás y la
tonalidad.
La transcripción polifónica es una tarea compleja que, hasta el momento,
no ha sido resuelta de manera eficaz para todos los tipos de sonidos armónicos.
Los mejores sistemas de detección de frecuencias fundamentales obtienen unos
porcentajes de acierto del 60%, aproximadamente. Se trata, principalmente,
de un problema de descomposición de señales en una mezcla, lo cual implica
conocimientos avanzados sobre procesamiento de señales digitales, aunque
debido a la naturaleza del problema también intervienen factores perceptuales,
psicoacústicos y musicológicos.
El proceso de transcripción puede separarse en dos tareas: convertir una
señal de audio en una representación de pianola, y convertir la pianola estimada
en notación musical.
Muchos autores sólo consideran la transcripción automática como una
conversión de audio a pianola, mientras que la conversión de pianola a notación
musical se suele ver como un problema distinto. La principal razón de esto es
que los procesos involucrados en la extracción de una pianola incluyen detección
de alturas y segmentación temporal de las notas, lo cual es una tarea ya de por
sı́ muy compleja. La conversión de pianola a partitura implica estimar el tempo,
cuantizar el ritmo o detectar la tonalidad. Esta fase está más relaccionada con
la generación de una notación legible para los músicos.

179
A. RESUMEN

En general, un sistema de transcripción musical no es capaz de obtener

con exactitud la partitura original que el músico ha interpretado. Las señales
musicales normalmente son interpretaciones expresivas, más que traducciones
literales de lo que puede leerse en una partitura. Por tanto, en una situación
real, una partitura se puede interpretar de muchas formas distintas, y trasladar
a una partitura las notas presentes en una señal de audio es un problema que
no tiene una única solución.
Sin embargo, la conversión de una señal musical en una representación de
pianola sin información rı́tmica ni armónica sólo depende de la información
contenida en la forma de la onda. Más que una representación orientada a
obtener partituras legibles, una pianola puede verse como una representación
orientada mostrar las frecuencias presentes en el sonido. La conversión de
una señal en una pianola se puede hacer mediante un método de detección
de frecuencias fundamentales. Este es el módulo principal de un sistema de
transcripción polifónica, ya que estima en número de notas presentes en cada
instante y sus alturas, inicios y duraciones. La mayorı́a de métodos para
estimación de frecuencia fundamental en señales polifónicas tienen un alto coste
computacional, debido a la complejidad de este problema.
La conversión de una pianola en una partitura legible requiere la extracción
de información armónica y rı́tmica. La tonalidad está relacionada con la
armonı́a, y las muestra relaciones de alturas en función de una tónica, que es el
centro de gravedad armónico. La separación de instrumentos y su clasificación
pueden usarse para identificar los distintos sonidos presentes en la señal,
permitiendo la extracción de partituras individuales para cada instrumento. La
estructura métrica se refiere a la organización temporal jerárquica, y especifica
cuántos tiempos hay en cada compás y cuál es la duración de nota usada para
representar un tiempo. De este modo, se pueden indicar los compases para
obtener una partitura legible. El tempo es una medida para especificar la
velocidad de ejecución de una pieza musical.
Cuando un músico interpreta una partitura, introduce desviaciones tempo-
rales, tanto intencionadas como involuntarias, y los inicios y duraciones de las
notas deben ajustarse (cuantizarse) para obtener una representación legible. La
presencia de estas desviaciones implica que, si se genera automáticamente un
sonido a partir de una pianola, este no coincididirá exactamente con los tiempos
de la partitura original. Esta es la razón principal por la que una pianola se
considera una representación puramente orientada al contenido acústico.
Esta tesis está principalmente enfocada a resolver el problema de estimación
de frecuencias fundamentales en señales polifónicas. Ésta es una tarea
extremadamente complicada de resolver y que ha sido abordada en numerosas
tesis doctorales.

180
 Las principales contribuciones de este trabajo son un conjunto de nuevos
métodos eficientes propuestos para la estimación de frecuencias fundamentales.
Estos algoritmos se han evaluado y comparado con otros métodos, dando buenos
resultados con un coste computacional muy bajo.
La detección de los comienzos de eventos musicales en señales de audio, o
detección de onsets, también se ha abordado en este trabajo, desarrollando un
método simple y eficiente para esta tarea. La información sobre onsets puede
usarse para estimar el tempo o para refinar un sistema de detección de alturas.
Los métodos propuestos se han aplicado a otras tareas de recuperación de
información musical, tales como clasificación de género, de modo, o identificación
de la autorı́a de una obra. Para ello, se ha combinado caracterı́sticas de audio
con caracterı́sticas simbólicas extraidas mediante la transcripción de señales
musicales, usando métodos de aprendizaje automático para obtener el género,
modo o autor.

2- Conocimientos previos8
Este capı́tulo introduce los conceptos necesarios para la adecuada comprensión
del trabajo. Se describen los conceptos y términos relacionados con métodos de
procesamiento de la señal, teorı́a musical y aprendizaje automático.
Primero, se hace una breve introducción de las distintas técnicas para el
análisis de señales de audio basadas en la transformada de Fourier, incluyendo
diferentes representaciones de tiempo-frecuencia.
A continuación, se analizan las propiedades de las señales musicales, y se
clasifican los instrumentos con respecto a su mecanismo de generación del sonido
y a sus caracterı́sticas espectrales.
También se abordan los conceptos necesarios sobre teorı́a musical, descri-
biendo las estructuras temporales y armónicas de la música occidental y su
representación usando notación escrita y computacional.
Finalmente, se describen las técnicas basadas en aprendizaje automático que
se han usado en este trabajo (redes neuronales y k vecinos más cercanos).

3 - Transcripción musical10
Este capı́tulo describe brevemente algunas caracterı́sticas perceptuales rela-
cionadas con el proceso que sigue un músico para realizar una transcripción
musical. Seguidamente, se analizan las limitaciones teóricas de la transcripción
automática desde un punto de vista del análisis y procesamiento de señales
discretas.
8 Background.
10 Music transcription.

181
A. RESUMEN

Finalmente, se introducen los conceptos básicos para la tarea de detección

de onsets. También se describen y analizan las distintas métricas usadas para
evaluar los sistemas de estimación de frecuencias fundamentales y detección de
onsets.

4 - Estado de la cuestión12

En este capı́tulo se presenta una descripción de los distintos sistemas previos

para estimación de una única frecuencia fundamental. Estos métodos se han
clasificado en los que analizan la forma de onda en el dominio del tiempo, los
que analizan la señal en el dominio de la frecuencia tras hacer una transformada
de Fourier, los métodos basados en modelos de percepción acústica y los modelos
probabilı́sticos.
Posteriormente, la revisión se extiende con una mayor cobertura a los
métodos de estimación de varias frecuencias fundamentales simultáneas. Es
complicado clasificar estos métodos usando una única taxonomı́a, ya que son
muy complejos y por tanto suelen incluir diferentes técnicas de procesamiento.
Por ejemplo, pueden categorizarse de acuerdo a su representación intermedia
(dominio del tiempo, transformada de Fourier de tiempo corto, wavelets, bancos
de filtros perceptuales, etc.), pero también respecto a su genericidad (algunos
métodos necesitan información a-priori sobre el instrumento a transcribir,
mientras que otros pueden usarse para analizar cualquier tipo de sonido
armónico), a su capacidad para modelar distintos timbres (por ejemplo, los
métodos paramétricos estadı́sticos pueden modelar envolventes cambiantes en
el tiempo o la frecuencia como aquellas producidas por un saxo, mientras que los
no paramétricos sólo pueden analizar patrones espectrales constantes, como los
que produce un piano), o por el modo en que pueden abordar las interacciones
entre distintos armónicos (métodos de estimación iterativa o conjunta).
En este trabajo, se ha propuesto una nueva categorización basada en la
metodologı́a principal que sigue el algoritmo, en lugar de la representación
intermedia escogida en la taxonomı́a de estimación de una única frecuencia
fundamental. Asimismo, también se han discutido y analizado los puntos fuertes
y débiles para cada una de estas categorı́as.
Finalmente, se ha hecho lo propio con los sistemas de detección de
onsets, clasificándolos en métodos de procesamiento de la señal y métodos de
aprendizaje automático.

12 State of the art.

182
 5 - Detección de onsets usando un banco de filtros
armónicos14
En este capı́tulo se propone un nuevo método para detección de onsets. La señal
de audio se analiza usando un banco de filtros pasa-banda de un semitono,
y se emplean las derivadas temporales de los valores filtrados para detectar
variaciones espectrales relacionadas con el inicio de los eventos musicales.
Este método se basa en las caracterı́sticas de los sonidos armónicos. Los
primeros cinco armónicos de un sonido afinado coinciden con las frecuencias de
otras notas en la afinación bien temperada usada en la música occidental. Otra
caracterı́stica importante de estos sonidos es que normalmente la mayor parte
de su energı́a se concentra en los primeros armónicos.
El banco de filtros de un semitono está formado por un conjunto de filtros
triangulares cuyas frecuencias centrales coinciden con las alturas musicales. En
la fase de sostenimiento y relajación de una nota, puede haber ligeras variaciones
en la intensidad y en la frecuencia de los armónicos. En este escenario, la
comparación espectral directa puede generar falsos positivos.
En cambio, usando el banco de filtros propuesto, se minimizan los efectos
de las variaciones espectrales sutiles que se producen durante las fases de
sostenimiento y relajación de una nota, mientras que en el ataque se incrementan
significativamente las amplitudes filtradas, ya que la mayor parte de energı́a de
los parciales se concentra en las frecuencias centrales de estas bandas. De este
modo, el sistema es especialmente sensible a variaciones de la frecuencia mayores
de un semitono, y por tanto se tiene en cuenta las propiedades armónicas de los
sonidos.
El método se ha evaluado y comparado con otros trabajos, dando buenos
resultados dada su sencillez, y obteniendo una alta eficiencia. El algoritmo,
desarrollado en C++, y la base de datos etiquetada para su evaluación se han
hecho públicos para futuras investigaciones.

6 - Estimación de alturas usando métodos de

aprendizaje supervisado16
En este capı́tulo se propone un método para la detección de alturas en piezas
musicales interpretadas por un solo instrumento con un patrón espectral simple.
Para ello, se parte de la hipótesis de que un paradigma de aprendizaje, tal
como una red neuronal, es capaz de inferir un patrón espectral tras una fase de
entrenamiento y, por tanto, detectar las notas en una pieza interpretada con el
14 Onset detection using a harmonic filter bank.
16 Multiple pitch estimation using supervised learning methods.

183
A. RESUMEN

mismo instrumento con el que se ha entrenado el sistema. Se ha puesto a prueba

dicha hipótesis, analizando su comportamiento y limitaciones con grabaciones
sintéticas.
Por tanto, la hipótesis de trabajo es que un método supervisado puede ser
capaz de aprender un patrón para un timbre determinado. Se ha entrenado
el sistema usando la información de las alturas correctas y espectrogramas
comprimidos mediante un banco de filtros de un semitono. Tras la fase de
aprendizaje, el algoritmo supervisado es capaz de detectar los patrones de las
alturas en el espectrograma, incluso en presencia de interferencias polifónicas
(solapamiento armónico). Para evaluar este método, se han comparado los
resultados usando redes neuronales dinámicas y k vecinos más cercanos.
El principal inconveniente de las técnicas de aprendizaje supervisado es que
dependen de los datos de entrenamiento. La cantidad de variedad tı́mbrica y
de combinaciones de distintas alturas pueden condicionar los resultados. Al ser
complicado conseguir datos musicales reales alineados con sus correspondientes
alturas, se han generado melodı́as sintetizadas en un escenario simplificado y
restringido, obteniendo resultados prometedores para sonidos sintéticos afinados
que poseen envolventes temporales constantes y patrones espectrales fijos.
Los resultados de la evaluación muestran que las redes neuronales son capaces
de identificar y generalizar los patrones espectrales obteniendo muy buenos
resultados para detección de alturas, mientras que los k vecinos no parecen
adecuados para esta tarea, dado el inmenso espacio de posibles prototipos
observables.

7 - Estimación de frecuencias fundamentales usando

métodos de procesamiento de la señal18
Los métodos de estimación de frecuencias fundamentales basados en aprendizaje
supervisado requieren datos de audio y simbólicos alineados para su entre-
namiento. Por tanto, estos métodos dependen del conjunto de entrenamiento,
y por este motivo muchos de ellos necesitan información a priori sobre el
timbre a analizar. Probablemente, es posible que estos sistemas puedan llegar a
generalizar e identificar correctamente las alturas en sonidos reales si se entrenan
usando un conjunto de datos suficientemente amplio, pero aun ası́ dependen de
los datos de entrenamiento.
En grabaciones reales puede haber varios instrumentos sonando simultánea-
mente, que son desconocidos a priori, y que además suelen presentar patrones
espectrales complejos. En este capı́tulo se describen tres métodos para detección
de frecuencias fundamentales en señales polifónicas que están completamente
18 Multiple fundamental frequency estimation using signal processing methods.

184
basados en métodos de procesamiento de la señal, evitando ası́ la necesidad de
emplear un conjunto de entrenamiento.
El primero de estos métodos es un algoritmo de cancelación iterativa,
principalmente enfocado a la transcripción de sonidos de cuerdas percutidas.
El principal objetivo de este sistema es obtener una estimación básica de las
frecuencias fundamentales presentes en señales reales, manteniendo un bajo
coste computacional. Este método se ha integrado con éxito en un sistema
más complejo para clasificación de géneros musicales.
Además del sistema de cancelación iterativa, se han propuesto dos nuevos
métodos de estimación conjunta que son capaces de tener en cuenta las
interacciones entre armónicos. Este tipo de métodos suele tener un alto coste
computacional debido a la evaluación de muchas posibles combinaciones de
alturas. Sin embargo, los métodos propuestos son muy eficientes. Estos métodos
se han evaluado y comparado con otros trabajos en MIREX 2007 y MIREX 2008,
obteniendo excelentes resultados con tiempos de ejecución muy bajos.
El primero de estos métodos realiza un análisis de la señal por ventanas,
obteniendo un conjunto de alturas en cada instante. Para ello, primero se
identifican los posibles candidatos a partir de los picos espectrales, y se generan
todas las combinaciones de candidatos para que un algoritmo de estimación
conjunta busque la mejor combinación teniendo en cuenta las interacciones entre
armónicos.
Para evaluar una combinación, se construye una secuencia hipotética de
parciales (HPS) para cada candidato. La puntuación de un candidato se
calcula considerando la suma de las amplitudes de los armónicos de su HPS
y una medida de suavidad de la envolvente espectral. La puntuación de una
combinación se calcula como la suma al cuadrado de las puntuaciones de sus
candidatos, y se selecciona la combinación de mayor puntuación en la ventana
actual.
El método asume que las envolventes espectrales de los sonidos analizados
tienden a variar suavemente en función de la frecuencia. El principio de suavidad
espectral se ha usado anteriormente (aunque de distinta manera) en trabajos
previos. La nueva medida de suavidad espectral está basada en la convolución
de la secuencia hipotética de parciales con una ventana gausiana.
Dada una combinación, la HPS de cada candidato se calcula teniendo en
cuenta las interacciones entre los armónicos de todos los candidatos de la
combinación. Para ello, primero se identifican los parciales solapados y se
estiman sus amplitudes por interpolación lineal usando las amplitudes de los
armónicos no solapados.
A diferencia del método de cancelación iterativa previamente descrito,
que asume un patrón armónico constante, el método de esimación conjunta
puede inferir patrones armónicos hipotéticos a partir de los datos espectrales,

185
A. RESUMEN

evaluándolos de acuerdo a las propiedades de los sonidos armónicos. Esta

metodologı́a es adecuada para la mayorı́a de los sonidos armónicos, a diferencia
del método de cancelación iterativa, el cual asume un patrón constante basado
en los sonidos de instrumentos de cuerda percutida.
En este método de esimación conjunta, cada ventana se analiza de manera
independiente, dando como resultado la combinación de frecuencias fundamen-
tales que maximiza una puntuación. Una de sus principales limitaciones es que
la información espectral que contiene una sola ventana se corresponde a un
periodo temporal breve y, debido a la naturaleza de las señales musicales, en
muchos casos es insuficiente para detectar las alturas de las notas, incluso para
músicos expertos.
Partiendo de la hipótesis de que el contexto temporal es importante, se ha
propuesto un segundo método de estimación conjunta, que extiende el anterior
considerando información de ventanas adyacentes para producir una detección
temporal suavizada. Adicionalmente, se ha incluı́do una técnica básica de
seguimiento de frecuencias fundamentales usando para ello un grafo acı́clico
dirigido, para considerar de este modo más información contextual.

8 - Conclusiones y trabajo futuro20

El objetivo principal de esta tesis es el desarrollo e implementación de un sistema
de transcripción polifónica. La investigación se centra principalmente en la
fase de detección de frecuencias fundamentales. Para lograr este objetivo, se
han analizado otros sistemas previos y se han propuesto distintas alternativas
basadas en el procesamiento de la señal de audio en el dominio de la frecuencia
y en su descomposición y segmentación. Se intenta que los métodos propuestos,
además de minimizar los errores, sean computacionalmente eficientes.
Para abordar el problema, se usan los datos de la señal en el dominio
de la frecuencia haciendo un análisis por ventanas mediante la transformada
de Fourier. Con estos datos de entrada se aplican distintas metodologı́as
para extraer las notas que están presentes en la señal, usando algoritmos de
aprendizaje (tales como vecinos más cercanos y redes neuronales dinámicas) y
desarrollando nuevos métodos basados en el procesamiento de señales. Para ello,
se tiene en cuenta la naturaleza armónica del sonido que genera la mayor parte
de los instrumentos musicales (la estructura de sus patrones espectrales), que
tratan de inferirse de la mezcla desarrollando técnicas de separación espectral y
de reconocimiento de patrones.
Las principales contribuciones de este trabajo pueden resumirse en los
siguientes puntos:
20 Conclusions and future work.

186
• Una revisión exhaustiva del estado de la cuestión para detección de
onsets y de frecuencias fundamentales en señales polifónicas. Los métodos
existentes se han clasificado en funciones prominentes (salience functions),
cancelación iterativa (iterative cancellation), estimación conjunta (joint
estimation), aprendizaje supervisado (supervised learning), aprendizaje no
supervisado (unsupervised learning), búsqueda de coincidencias (matching
pursuit), modelos bayesianos (Bayesian models), modelos espectrales
estadı́sticos (statistical spectral models), sistemas de pizarra (blackboard
systems), y métodos de coincidencia con bases de datos (database
matching). Se ha hecho un análisis de los puntos fuertes y débiles de
cada una de estas categorı́as.

• El desarrollo de un sistema eficiente para la detección de onsets y la

construcción de un conjunto de datos etiquetado para esta tarea. Las
principales novedades en este campo son el uso de un banco de filtros
de un doceavo de octava para comprimir la información armónica y
las sencillas funciones de detección de onsets propuestas. El método
presentado está principalmente indicado para la detección de onsets
percusivos, ya que detecta cambios bruscos de energı́a, pero también
considera las propiedades de las señales armónicas, lo cual hace que el
sistema sea robusto ante variaciones espectrales producidas durante la
fase de sostenimiento de los sonidos. El algoritmo ha sido evaluado y
comparado con otros trabajos, obteniendo resultados satisfactorios.

• Dos nuevos métodos para la estimación de frecuencias fundamentales de

instrumentos conocidos a priori usando técnicas de aprendizaje super-
visado. Estos algoritmos fueron unos de los primeros métodos basados en
aprendizaje automático que se han propuesto para esta tarea. Un banco
de filtros armónicos se ha usado para reducir la cantidad de información
espectral que se usa como entrada para una red neuronal de tiempo
retardado preservando el principal contenido armónico. Se ha generado
un conjunto de entrenamiento y validación para evaluar el método.
Las conclusiones extraı́das tras comparar para esta tarea los resultados
obtenidos por los k vecinos más cercanos y las redes neuronales de tiempo
retardado también son relevantes. La red neuronal claramente mejora
los resultados obtenidos por los vecinos más cercanos usando sonidos
sintetizados, mostrando las ventajas de las redes para la generalización en
un espacio de observaciones muy extenso. Se han propuesto funciones de
activación alternativas para generalizar los prototipos obtenidos mediante
vecinos más cercanos, pero los resultados han seguido siendo claramente
inferiores a los obtenidos con la red neuronal.

187
A. RESUMEN

• Un método sencillo de cancelación iterativa, principalmente orientado a la

transcripción de música de piano con un coste computacional muy bajo.
Se ha desarrollado un sistema completo para cargar un fichero de audio
y que obtiene un fichero MIDI como resultado. Se ha propuesto una
nueva arquitectura, basada en el análisis aislado de aquellos espectros
que están tras cada onset detectado (para aumentar la eficiencia), y se
ha evaluado la aplicación de un método para la extracción de sinusoides
en esta tarea. El sistema propuesto ha sido la base para los métodos
de estimación conjunta desarrollados con posterioridad, y ha sido usado
con éxito para clasificación de género y otras tareas de recuperación de
información musical.

• Métodos de estimación de frecuencias fundamentales en señales polifónicas

basados en técnicas de procesamiento de señal para analizar música real
sin ningún conocimiento a priori. Estos métodos, que son probablemente
la mayor contribución de esta tesis, han alcanzado experimentalmente a
los mejores algoritmos para esta tarea con un alto grado de eficiencia.
Los patrones armónicos se han inferido de la información espectral,
teniendo en cuenta la intensidad y una métrica propuesta para medir
la suavidad de la envolvente en el dominio de la frecuencia, en un
esquema de evaluación conjunta que tiene en cuenta las interacciones
entre las distintas fuentes. Estos métodos también introducen un proceso
de suavizado temporal considerando las combinaciones de alturas en
ventanas adyacentes. También se ha propuesto un sistema sencillo de
seguimiento de frecuencias fundamentales usando un grafo dirigido acı́clico
y ponderado. Estos nuevos métodos han conseguido tasas de acierto muy
altas, aunque también poseen una alta eficiencia, que era la mayor barrera
a la que se enfrentaban los métodos existentes de estimación conjunta.

Lineas de trabajo futuro

Como trabajo futuro, los métodos propuestos basados en redes neuronales
podrı́an adaptarse facilmente para detección de onsets. Al igual que en los
sistemas de aprendizaje automático descritos para detección de frecuencias
fundamentales, el banco de filtros de un doceavo de octava podrı́a usarse para
obtener los datos de entrada a la red, peo en el caso de esta tarea sólo una
neurona serı́a necesaria en la capa de salida para clasificar cada ventana como
onset o no onset. Este esquema de aprendizaje también implica ampliar la base
de datos compilada para detección de onsets para obtener un conjunto de datos
mayor y, por tanto, más fiable.
Los métodos de aprendizaje supervisado también podrı́an ser entrenados y
evaluados usando mezclas de distintos instrumentos reales alineados con sus

188
correspondientes alturas etiquetadas. Como se ha descrito en este trabajo, no
es una tarea sencilla alinear una base de datos de este tipo. Sin embargo, esta
es una linea de investigación que deberı́a ser explorada.
Probablemente, los sistemas de estimación de frecuencias fundamentales
que sólo consideran ventanas individuales no podrán ser capaces de mejorar
significativamente los resultados actuales. La cantidad de datos presente en el
periodo correspondiente a una ventana no es suficiente para detectar las alturas,
incluso para un músico experto. El contexto juega un papel muy importante
en la música. Por ejemplo, es muy complicado identificar las alturas cuando
escuchamos dos canciones que no están sincronizadas sonando simultáneamente,
incluso si no son muy complejas.
Por tanto, la tarea de estimación de las alturas deberı́a complementarse
de algún modo con información temporal. La coherencia de las detecciones
a lo largo del tiempo se ha considerado en uno de los métodos propuestos,
pero podrı́a extenderse usando un sistema fiable de seguimiento de frecuencias
fundamentales. Sin embargo, esta tarea es complicada desde un punto de vista
computacional.
La investigación futura de las arquitecturas alternativas propuestas para los
métodos de estimación conjunta también es una linea de trabajo prometedora.
La metodologı́a de este sistema permite, por ejemplo, analizar conjuntamente
combinaciones de aquellas ventanas que están entre dos onsets consecutivos,
obteniendo las alturas para este intervalo. Perceptualmente, los resultados
obtenidos con este esquema fueron mejores que analizando aisladamente con-
juntos adyacentes de ventanas. Sin embargo, la menor resolución temporal y los
errores en el método de detección de onsets, sumado al problema de detección de
los offsets en el intervalo entre dos onsets, condicionan el porcentaje de acierto
usando una métrica de evaluación clásica. Como trabajo futuro, está planeado
evaluar estas arquitecturas usando una métrica perceptual.
La información multimodal también es una linea de trabajo prometedora. La
inclusión de modelos musicales considerando la tonalidad, el tempo o la métrica
para inferir probabilidades de notas podrı́a complementar las estimaciones de
las alturas. Estas lineas de trabajo están planeadas dentro del proyecto DRIMS
en colaboración con el Music Technology Group de la Universitat Pompeu Fabra
de Barcelona.
La transcripción musical interactiva también está planeada como trabajo
futuro dentro del proyecto MIPRCV (Consolider Ingenio 2010). Se trata
de desarrollar un método asistido por ordenador para transcripción musical.
Usando un interfaz visual, los segmentos transcritos automáticamente pueden
ser aceptados o corregidos por un músico experto. Después, estos segmentos
validados pueden usarse como información para el método de estimación de

189
A. RESUMEN

frecuencias fundamentales, produciendo ası́ sugerencias mejoradas y corrigiendo

errores frecuentes.
La investigación futura de los métodos propuestos aplicados a otras tareas
de recuperación de información musical, tales como clasificación de género o
identificación de autor, también se está desarrollando en colaboración con el
centro de investigación de Software Technology and Interactive Systems (Vienna
University of Technology).

190
 Bibliography

Abdallah, S. A. and Plumbley, M. D. (2003a). An ICA approach to automatic

music transcription. In Proc. 114th AES Convention. (Cited on page 70).

Abdallah, S. A. and Plumbley, M. D. (2003b). Probability as metadata: Event

detection In music using ICA as a conditional density model. In Proc. of the
Fourth International Symposium on Independent Component Analysis (ICA),
pages 233–238, Nara, Japan. (Cited on page 81).

Abdallah, S. A. and Plumbley, M. D. (2004). Polyphonic music transcription

by non-negative sparse coding of power spectra. In Proc. of the 5th
International Conference on Music Information Retrieval (ISMIR), pages
318–325, Barcelona, Spain. (Cited on pages 70 and 77).

Ahmed, N., Natarjan, T., and Rao, K. (1974). Discrete cosine transform. IEEE
Trans. on Computers, 23:90–93. (Cited on page 16).

American Standards Association (1960). American standard acoustical

terminology. Definition 12.9. Timbre. (Cited on page 19).

Bay, M., Ehmann, A. F., and Downie, J. S. (2009). Evaluation of multiple-

f0 estimation and tracking systems. In Proc. of the 10th International
Conference on Music Information Retrieval (ISMIR), pages 315–320. (Cited
on pages xiii, 157, 162, 165, 166, and 167).

Beauchamp, J. W., Maher, R. C., and Brown, R. (1993). Detection of musical

pitch from recorded solo performances. In Proc. 1993 Audio Engineering
Society Convention, pages 1–15, Berlin, Germany. Preprint 3541. (Cited on
page 48).

Bello, J. P. (2000). Blackboard system and top-down processing for the

transcription of simple polyphonic music. In Proc. of the COST G-6
Conference on Digital Audio Effects (DAFx), Verona, Italy. (Cited on pages
74, 75, and 77).

Bello, J. P. (2004). Towards the Automated Analysis of Simple Polyphonic

Music: A Knowledge-based Approach. PhD thesis, University of London, UK.
(Cited on page 4).

Bello, J. P., Daudet, L., Abdallah, S., Duxbury, C., Davies, M., and Sandler,
M. B. (2005). A tutorial on onset detection in music signals. IEEE Trans. on
Speech and Audio Processing, 13(5):1035–1047. (Cited on pages 52, 53, 77,
and 78).

191
BIBLIOGRAPHY

Bello, J. P., Daudet, L., and Sandler, M. (2002). Time-domain polyphonic

transcription using self-generating databases. In Proc. of the 112th
Convention of the Audio Engineering Society, Munich, Germany. (Cited
on pages 75 and 123).

Bello, J. P., Duxbury, C., Davies, M., and Sandler, M. (2004). On the use of
phase and energy for musical onset detection in the complex domain. IEEE
Signal Processing Letters, 11(6):553–556. (Cited on page 79).

Bello, J. P. and Sandler, M. (2003). Phase-based note onset detection for

music signals. In International Conference on Acoustics, Speech and Signal
Processing (ICASSP), volume V, pages 441–444, Hong Kong. (Cited on page
79).

Bertin, N., Badeau, R., and Richard, G. (2007). Blind signal decompositions
for automatic transcription of polyphonic music: NMF and K-SVD on the
benchmark. In IEEE Int. Conf. on Acoustics, Speech and Signal Processing
(ICASSP), volume I, pages 65–68, Honolulu, HI. (Cited on page 70).

Bilmes, J. (1993). Timing is of the Essence: Perceptual and Computational

Techniques for Representing, Learning and Reproducing Expressive Timing
in Percusive Rhythm. MSc Thesis, MIT. (Cited on pages 32 and 85).

Boulanger, R. (1999). The CSound Book. MIT Press, Cambridge,

Massachusetts. (Cited on page 37).

Bourlard, H. A. and Morgan, N. (1992). Connectionist speech recognition. A

Hybrid approach. Kluwer Academic Publishers. (Cited on page 102).

Bregman, A. S. (1990). Auditory Scene Analysis. MIT Press, Cambridge, MA.

(Cited on page 44).

Brossier, P. (2005). Fast onset detection using aubio. In MIREX (2005), onset
detection contest. (Cited on page 92).

Brossier, P., Bello, P., and Plumbley, D. (2004). Real-time temporal

segmentation of note objects in music signals. In Proc. of the International
Computer Music Conference (ICMC), Florida. (Cited on pages 78 and 79).

Brown, J. C. (1991). Calculation of a constant Q spectral transform. Journal

of the Acoustical Society of America, 89(1):425–434. (Cited on page 14).

Brown, J. C. (1992). Musical fundamental frequency tracking using a pattern

recognition method. Journal of the Acoustical Society of America, 92(3):1394–
1402. (Cited on page 58).

192
 BIBLIOGRAPHY

Brown, J. C. and Puckette, M. S. (1992). An efficient algorithm for the

calculation of a constant Q transform. Journal of the Acoustical Society of
America, 92(5):2698–2701. (Cited on page 15).

Burges, C. J. C. (1998). A tutorial on support vector machines for pattern

recognition. In Data Mining and Knowledge Discovery, pages 121–167. Kluwe
Academic Publishers, Boston. (Cited on page 69).

Butterworth, S. (1930). On the Theory of Filter Amplifiers. Wireless Engineer,

7:536–541. (Cited on page 94).

Cambouropoulos, E. (2000). From MIDI to traditional musical notation. In

Proc. of the AAAI 2000 Workshop on Artificial Intelligence and Music:
Towards Formal Models for Composition, Performance and Analysis. 17th
National Conference on Artificial Intelligence (AAAI), Austin, TX. (Cited
on page 37).

Cano, P. (1998). Fundamental frequency estimation in the SMS analysis. In

Proceedings of the Digital Audio Effects Workshop (DAFx). (Cited on page
59).

Cao, C. and Li, M. (2008). Multiple F0 estimation in polyphonic music (MIREX

2008). In MIREX (2008), multiple f0 estimation and tracking contest. (Cited
on page 163).

Cao, C., Li, M., Liu, J., and Yan, Y. (2007). Multiple f0 estimation in polyphonic
music. In MIREX (2007), multiple f0 estimation and tracking contest. (Cited
on pages 65 and 160).

Carpenter, G. A., Grossberg, S., and Reynolds, J. (1991). ARTMAP: Supervised

real-time learning and classification of nonstationary data by a self-organizing
neural network. Neural Networks, 4:493–504. (Cited on page 102).

Cañadas-Quesada, F. J., Vera-Candeas, P., Ruiz-Reyes, N., and Carabias-Orti,

J. J. (2009). Polyphonic transcription based on temporal evolution of spectral
similarity of gaussian mixture models. In 17th European Signal Processing
Conference (EUSIPCO), pages 10–14, Glasgow, Scotland. (Cited on page
66).

Cañadas-Quesada, F. J., Vera-Candeas, P., Ruiz-Reyes, N., Mata-Campos, R.,

and Carabias-Orti, J. J. (2008). Note-event detection in polyphonic musical
signals based on harmonic matching pursuit and spectral smoothness. Journal
of New Music Research, 37(3):167–183. (Cited on pages 71 and 127).

Cemgil, A. T. (2004). Bayesian Music Transcription. PhD thesis, Radboud

University of Nijmegen, Netherlands. (Cited on pages 4 and 72).

193
BIBLIOGRAPHY

Cemgil, A. T., Kappen, B., and Barber, D. (2003). Generative model based
polyphonic music transcription. In IEEE Workshop on Applications of Signal
Processing to Audio and Acoustics, pages 181–184. (Cited on pages 1 and 72).

Cemgil, A. T., Kappen, H. J., and Barber, D. (2006). A generative model for
music transcription. IEEE Trans. on Audio, Speech and Language Processing,
14(2):679–694. (Cited on page 72).

Chang, W. C., Su, A. W. Y., Yeh, C., Roebel, A., and Rodet, X. (2008).
Multiple-F0 tracking based on a high-order HMM model. In Proc. of the 11th
Int. Conference on Digital Audio Effects (DAFx), Espoo, Finland. (Cited on
pages 66, 168, and 173).

Cohen, L. (1995). Time-frequency analysis. Prentice Hall. (Cited on page 66).

Collins, N. (2005a). A change discrimination onset detector with peak scoring

peak picker and time domain correction. In MIREX (2005), onset detection
contest. (Cited on page 99).

Collins, N. (2005b). A comparison of sound onset detection algorithms with

emphasis on psychoacoustically motivated detection functions. In AES
Convention 118, pages 28–31, Barcelona. (Cited on pages 77 and 79).

Comon, P. (1994). Independent component analysis, a new concept? Signal

processing, 36:287–314. (Cited on page 70).

Cont, A. (2006). Realtime multiple pitch observation using sparse non-

negative constraints. In Proc. of the 7th International Symposium on Music
Information Retrieval (ISMIR), Victoria, Canada. (Cited on page 69).

Cont, A. (2007). Real-time transcription of music signals: MIREX 2007

submission description. In MIREX (2007), multiple f0 estimation and
tracking contest. (Cited on pages 159 and 160).

Cont, A. (2008). Modeling Musical Anticipation: From the time of music to the
music of time. PhD thesis, University of Paris VI and University of California
in San Diego. (Cited on page 44).

Cover, T. and Hart, P. (1967). Nearest neighbor pattern classification. IEEE

Trans. on Information Theory, 13(1):21 – 27. (Cited on page 40).

Daniel, A., Emiya, V., and David, B. (2008). Perceptually-based evaluation

of the errors usually made when automatically transcribing music. In Proc.
of the 9th Int. Conference on Music Information Retrieval (ISMIR), pages
550–555, Philadelphia, PA. (Cited on page 52).

194
 BIBLIOGRAPHY

Daubechies, I. (1988). Orthonormal bases of compactly supported wavelets.

Communications on Pure and Applied Mathematics, 41(7):909–996. (Cited
on page 13).

Daubechies, I. (1992). Ten Lectures on Wavelets. Society for Industrial &

Applied Mathematics. (Cited on page 13).

Daudet, L. (2001). Transients modelling by pruned wavelet tress. In Proc. of

the Int. Computer Music Conference (ICMC), pages 18–21. (Cited on page
78).

Davy, M. (2006a). An introduction to signal processing. In Klapuri and Davy

(2006), chapter 2. (Cited on page 58).

Davy, M. (2006b). Multiple fundamental frequency estimation based on

generative models. In Klapuri and Davy (2006), chapter 7. (Cited on page
72).

Davy, M. and Godsill, S. (2002). Detection of abrupt spectral changes

using support vector machines. an application to audio signal segmentation.
In International Conference on Acoustics, Speech and Signal Processing
(ICASSP), pages 1313–1316. (Cited on page 81).

Davy, M., Godsill, S. J., and Idier, J. (2006). Bayesian analysis of polyphonic
western tonal music. Journal of the Acoustical Society of America, 119:2498–
2517. (Cited on page 72).

de Cheveigné, A. (1993). Separation of concurrent harmonic sounds:

Fundamental frequency estimation and a time-domain cancellation model for
auditory processing. Journal of the Acoustical Society of America, 93(6):3271–
3290. (Cited on page 65).

de Cheveigné, A. (2005). Pitch perception models. In Plack, C. J., Oxenham,

A. J., Fay, R. R., and Popper, A. N., editors, Pitch: neural coding and
perception, chapter 6. Springer. (Cited on pages 60 and 65).

de Cheveigné, A. and Kawahara, H. (2002). YIN, a fundamental frequency

estimator for speech and music. Journal of the Acoustical Society of America,
111(4):1917–1930. (Cited on pages 56 and 173).

Deutsch, D. (1998). The psychology of music, 2nd edition (cognition and

perception). Academic press. (Cited on pages 26 and 198).

Dijkstra, E. W. (1959). A note on two problems in connexion with graphs.

Numerische Mathematik, pages 269–271. (Cited on page 139).

195
BIBLIOGRAPHY

Dixon, S. (2006). Onset detection revisited. In Proc. of the Int. Conf. on Digital
Audio Effects (DAFx), pages 133–137, Montreal, Canada. (Cited on pages
77, 78, 79, 91, 92, and 140).

Doval, B. (1994). Estimation de la Fréquence Fondamentale des signaux sonores.

PhD thesis, Université Paris VI, Paris. (Cited on page 121).

Doval, B. and Rodet, X. (1993). Fundamental frequency estimation and tracking

using maximum likelihood harmonic matching and HMMs. In International
Conference on Acoustics, Speech and Signal Processing (ICASSP), volume 1,
pages 221–224. (Cited on page 58).

Dubnowski, J. J., Schafer, R. W., and Rabiner, L. R. (1976). Real-time

digital hardware pitch detector. IEEE Trans. Acoustics, Speech, and Signal
Processing (ASSP), 24:2–8. (Cited on page 55).

Dubois, C. and Davy, M. (2005). Harmonic tracking using sequential Monte

Carlo. In IEEE/SP 13th Workshop on Statistical Signal Processing, pages
1292–1296, Bordeaux, France. (Cited on page 72).

Dubois, C. and Davy, M. (2007). Joint Detection and Tracking of Time-Varying

Harmonic Components: A Flexible Bayesian Approach. IEEE Trans. on
Audio, Speech and Language Processing, 15(4):1283–1295. (Cited on page
72).

Duda, R., Lyon, R., and Slaney, M. (1990). Correlograms and the separation
of sounds. In Proc. IEEE Asilomar Conference on Signals, Systems and
Computers. (Cited on page 60).

Duda, R. O., Hart, P. E., and Stork, D. G. (2000). Pattern Classification. John
Wiley and Sons. (Cited on pages xi, 38, 39, 40, 41, and 104).

Durrieu, J. L., Richard, G., and David, B. (2008). Singer melody extraction
in polyphonic signals using source separation methods. In Proc of the
IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), pages 169–172, Las Vegas, NV. (Cited on page 163).

Duxbury, C., Sandler, M., and Davies, M. (2002). A hybrid approach to musical
note onset detection. In Proc. Digital Audio Effects Conference (DAFx), pages
33–38, Hamburg, Germany. (Cited on pages 78, 80, and 83).

Egashira, K., Ono, N., and Sagayama, S. (2008). Sequential estimation

of multiple fundamental frequencies through harmonic-temporal-structured
clustering. In MIREX (2008), multiple f0 estimation and tracking contest.
(Cited on pages 163 and 164).

196
 BIBLIOGRAPHY

Ellis, D. P. W. (1996). Prediction-driven computational auditory scene analysis.

PhD thesis, MIT. (Cited on page 74).

Emiya, V. (2008). Automatic transcription of piano music. PhD thesis, Ecole

Nationale Supérieure des Télécommunications (ENST), Paris, France. (Cited
on page 4).

Emiya, V., Badeau, R., and David, B. (2007). Multipitch estimation and
tracking of inharmonic sounds in colored noise. In MIREX (2007), multiple
f0 estimation and tracking contest. (Cited on pages 159 and 160).

Emiya, V., Badeau, R., and David, B. (2008a). Automatic transcription of

piano music based on HMM tracking of jointly-estimated pitches. In MIREX
(2008), multiple f0 estimation and tracking contest. (Cited on pages 163
and 164).

Emiya, V., Badeu, R., and David, B. (2008b). Automatic transcription of piano
music based on HMM tracking of jointly-estimated pitches. In Proc. European
Signal Processing Conference (EUSIPCO), Rhodes, Greece. (Cited on pages
66, 123, 124, 128, 129, 159, and 160).

Engelmore, R. S. and Morgan, A. J. (1988). Blackboard Systems. Addison-

Wesley publishing. (Cited on page 74).

Every, M. R. and Szymanski, J. E. (2006). Separation of synchronous pitched

notes by spectral filtering of harmonics. IEEE Trans. on Audio, Speech, and
Language Processing, 14(5):1845–1856. (Cited on pages 123 and 130).

FitzGerald, D. (2004). Automatic drum transcription and source separation.

PhD thesis, Dublin Inst. Technol. (Cited on pages 26 and 70).

Fitzgerald, D. and Paulus, J. (2006). Unpitched percussion transcription. In

Klapuri and Davy (2006), chapter 5. (Cited on page 26).

Fletcher, H. and Munson, W. A. (1933). Loudness, its definition, measurement

and calculation. Journal of the Acoustical Society of America, 5:82–108.
(Cited on page 10).

Fletcher, N. H. and Rossing, T. D. (1988). The physics of musical instruments.

Springer, Berlin. (Cited on pages 23 and 124).

Fonseca, N. and Ferreira, A. (2009). Measuring music transcription results based

on a hybrid decay/sustain evaluation. In Proc. of the 7th Triennial Conference
of European Society for the Cognitive Sciences of Music (ESCOM), pages
119–124, Jyväskylä, Finland. (Cited on page 52).

197
BIBLIOGRAPHY

Fraisse, P. (1998). Rhythm and tempo. In Deutsch (1998), chapter 6. (Cited

on page 31).

Frigo, M. and Johnson, S. G. (2005). The design and implementation of

FFTW3. Proceedings of the IEEE, 93(2):216–231. Special issue on “Program
Generation, Optimization, and Platform Adaptation”. (Cited on pages 9
and 120).

Gabor, D. (1946). Theory of communication. J. Inst. Elect. Eng., 93:429–457.

(Cited on page 71).

Gabor, D. (1947). Acoustical quanta and the theory of hearing. Nature,

159(4044):591–594. (Cited on page 71).

Gerhard, D. (2003). Pitch extraction and fundamental frequency: History and

current techniques. Technical report, University of Regina. (Cited on page
55).

Goto, M. (2000). A robust predominant f0 estimation method for real-

time detection of melody and bass lines in cd recordings. In Proc. of the
IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), volume II, pages 757–760, Istambul, Turkey. (Cited on pages xii,
72, and 73).

Goto, M. (2003). RWC (Real World Computing) Music database. http://

staff.aist.go.jp/m.goto/RWC-MDB/. (Cited on pages 10, 18, 21, 22, 25,
27, 38, 86, 88, 89, 90, and 157).

Goto, M. and Muraoka, Y. (1995). A real-time beat tracking system for audio
signals. In Proc. of International Computer Music Conference (ICMC), pages
171–174. (Cited on page 85).

Goto, M. and Muraoka, Y. (1996). Beat tracking based on multiple-agent

architecture — a real-time beat tracking system for audio signals —. In Proc.
of the Second Int. Conf. on Multi-Agent Systems, pages 103–110. (Cited on
page 85).

Gouyon, F. (2008). Computational rhythm description. A review and novel

approach. VDM Verlag Dr Müller. (Cited on page 31).

Grey, J. (1978). Timbre discrimination in musical patterns. Journal of the

Acoustical Society of America, 64:467–472. (Cited on page 19).

Gribonval, R. and Bacry, E. (2003). Harmonic decomposition of audio signals

with matching pursuit. IEEE Trans. on Signal Processing, 51(1):101–111.
(Cited on page 71).

198
 BIBLIOGRAPHY

Griffin, D. W. and Lim, J. S. (1985). A new model-based speech

analysis/synthesis system. In Proc. of the IEEE International Conference
on Acoustics, Speech, and Signal Processing (ICASSP), volume 10, pages
513–516, Tampa, FL. (Cited on page 121).

Groble, M. (2008). Multiple fundamental frequency estimation. In MIREX

(2008), multiple f0 estimation and tracking contest. (Cited on pages 75
and 163).

Haar, A. (1911). Zur theorie der orthogonalen funktionensysteme.

Mathematische Annalen, 71(1):38–53. (Cited on page 13).

Hainsworth, S. (2003). Techniques for the Automated Analysis of Musical Audio.

PhD thesis, Signal Processing Group, Department of Engineering, University
of Cambridge. (Cited on pages xi, 4, 32, 44, and 76).

Handel, S. (1989). Listening: An introduction to the perception of auditory

events. Bradford Books / MIT Press. (Cited on page 31).

Harris, F. J. (1978). On the use of windows for harmonic analysis with the
discrete Fourier transform. Proceedings of the IEEE, 66(1):51–83. (Cited on
page 11).

Hartmann, W. M. (1996). Pitch, periodicity, and auditory organization. Journal

of the Acoustical Society of America, 100(6):3491–3502. (Cited on page 21).

Herrera, P., Klapuri, A., and Davy, M. (2006). Automatic classification of

pitched musical instrument sounds. In Klapuri and Davy (2006), chapter 6.
(Cited on page 20).

Hess, W. J. (1983). Algorithms and Devices for Pitch Determination of Speech-

Signals. Springer-Verlag, Berlin. (Cited on page 55).

Holland, J. H. (1992). Adaptation in Natural and Artificial Systems: An

Introductory Analysis with Applications to Biology, Control, and Artificial
Intelligence. MIT Press. (Cited on page 68).

Honing, H. (2001). From time to time: The representation of timing and tempo.
Computer Music Journal, 25(3):50–61. (Cited on page 31).

Hornbostel, E. M. and Sachs, C. (1914). Systematik der musikinstrumente.

Zeitschrift für Ethnologie, 46:553–590. (Cited on pages 20 and 24).

Huang, X., Acero, A., and Hon, H. (2001). Spoken Language Processing: A
guide to theory, algorithm, and system development. Prentice Hall. (Cited
on pages 16 and 22).

199
BIBLIOGRAPHY

Huron, D. (1989). Voice denumerability in polyphonic music of homogeneous

timbres. Music Perception, 6(4):361–382. (Cited on page 51).

Hush, D. R. and Horne, B. G. (1993). Progress in supervised neural networks.

IEEE Signal Processing Magazine, 1(10):8–39. (Cited on page 39).

Jensen, K. (1999). Envelope model of isolated musical sounds. In Proc. Digital

Audio Effects Conference (DAFx). (Cited on page 18).

Juslin, P. N., Karlsson, J., Lindström, E., Friberg, A., and Schoonderwaldt,
E. (2006). Play it again with feeling: Computer feedback in musical
communication of emotions. Journal of Experimental Psychology: Applied,
12(2):79–95. (Cited on page 33).

Kameoka, H., Nishimoto, T., and Sagayama, S. (2007). A multipitch analyser

based on harmonic temporal structured clustering. IEEE Trans. on Audio,
Speech and Language Processing, 5(3):982–994. (Cited on pages xii, 73, 74,
159, and 160).

Kapanci, E. and Pfeffer, A. (2004). A hierarchical approach to onset detection.

In In Proc. International Computer Music Conference (ICMC), pages 438–
441. (Cited on page 81).

Kashino, K. and Godsill, S. J. (2004). Bayesian estimation of simultaneous

musical notes based on frequency domain modeling. In Int. Conference
on Acoustics, Speech, and Signal Processing (ICASSP), Montreal, Canada.
(Cited on page 72).

Kashino, K. and Tanaka, H. (1993). A sound source separation system with

the ability of automatic tone modeling. In Proc. of International Computer
Music Conference (ICMC), pages 248–255. (Cited on page 72).

Klapuri, A. (1998). Number theoretical means of resolving a mixture of

several harmonic sounds. In Proc. European Signal Processing Conference
(EUSIPCO), Rhodes, Greece. (Cited on page 46).

Klapuri, A. (1999). Sound onset detection by applying psychoacoustic

knowledge. In IEEE Int. Conf. on Acoustics, Speech and Signal Processing
(ICASSP), pages 3089–3092, Phoenix, USA. (Cited on pages 77, 78, 79,
and 83).

Klapuri, A. (2001). Multipitch estimation and sound separation by the spectral

smoothness principle. In IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pages 3381–3384, Salt Lake City, Utah.
(Cited on page 63).

200
 BIBLIOGRAPHY

Klapuri, A. (2003a). Multiple fundamental frequency estimation based on

harmonicity and spectral smoothness. IEEE Trans. on Speech and Audio
Processing, 11(6):804–816. (Cited on pages xi, 47, 60, 63, 71, 120, 127,
and 130).

Klapuri, A. (2003b). Musical meter estimation and music transcription. In Proc.

Cambridge Music Processing Colloquium, pages 40–45. (Cited on pages 31
and 32).

Klapuri, A. (2004). Signal processing methods for the automatic transcription

of music. PhD thesis, Tampere Univ. of Technology. (Cited on pages 4, 46,
62, 76, and 128).

Klapuri, A. (2005). A perceptually motivated multiple-f0 estimation method

for polyphonic music analysis. In IEEE workshop on Applications of Signal
Processing to Audio and Acoustics (WASPAA), New Paltz, NY. (Cited on
page 64).

Klapuri, A. (2006a). Introduction to music transcription. In Klapuri and Davy

(2006), chapter 1. (Cited on page 21).

Klapuri, A. (2006b). Multiple fundamental frequency estimation by summing

harmonic amplitudes. In Proc. of the Int. Conference on Music Information
Retrieval (ISMIR), pages 216–221, Victoria, Canada. (Cited on pages 64,
120, 125, 141, and 149).

Klapuri, A. (2008). Multipitch analysis of polyphonic music and speech

signals using an auditory model. IEEE Trans. Audio, Speech and Language
Processing, 16(2):255–266. (Cited on page 64).

Klapuri, A. and Astola, J. (2002). Efficient calculation of a physiologically-

motivated representation for sound. In Proceedings of IEEE International
Conference on Digital Signal Processing. (Cited on page 60).

Klapuri, A. and Davy, M. (2006). Signal processing methods for music

transcription. Springer. (Cited on pages 37, 195, 197, 199, 201, 209, and 212).

Klapuri, A., Eronen, A. J., and Astola, J. T. (2006). Analysis of the meter
of acoustic musical signals. IEEE Trans. on Audio, Speech and Language
Processing, 14(1):342–355. (Cited on page 32).

Klapuri, A., Virtanen, T., and Holm, J.-M. (2000). Robust multipitch
estimation for the analysis and manipulation of polyphonic musical signals. In
Proc. COST-G6 Conference on Digital Audio Effects (DAFx), pages 233–236.
(Cited on page 44).

201
BIBLIOGRAPHY

Kosuke, I., Ken’Ichi, M., and Tsutomu, N. (2003). Ear advantage and
consonance of dichotic pitch intervals in absolute-pitch possessors. Brain and
cognition, 53(3):464–471. (Cited on page 31).

Krstulovic, S. and Gribonval, R. (2006). MPTK: Matching Pursuit made

tractable. In Proc. Int. Conf. on Acoustics, Speech, and Signal Processing
(ICASSP), volume III, pages 496–499, Toulouse, France. (Cited on page 70).

Krumhansl, C. (2004). The cognition of tonality - as we know it today. Journal

of New Music Research, 33(3):253–268. (Cited on pages xi, 30, and 31).

Lacoste, A. and Eck, D. (2005). Onset detection with artificial neural networks.
In MIREX (2005), onset detection contest. (Cited on page 78).

Lacoste, A. and Eck, D. (2007). A supervised classification algorithm for note

onset detection. EURASIP Journal on Advances in Signal Processing. (Cited
on page 81).

Lahat, M., Niederjohn, R., and Krubsack, D. (1987). A spectral autocorrelation

method for measurement of the fundamental frequency of noise-corrupted
speech. IEEE Trans. on Acoustics, Speech and Signal Processing, 35(6):741 –
750. (Cited on page 57).

Large, E. W. and Kolen, J. F. (1994). Resonance and the perception of musical

meter. Connection Science, 6:279–312. (Cited on page 67).

Lee, D. D. and Seung, H. (1999). Learning the parts of objects by non-negative

matrix factorization. Nature, 401:788–791. (Cited on page 69).

Lee, W.-C. and Kuo, C.-C. J. (2006). Musical onset detection based on adaptive
linear prediction. IEEE International Conference on Multimedia and Expo,
0:957–960. (Cited on page 78).

Lee, W.-C., Shiu, Y., and Kuo, C.-C. J. (2007). Musical onset detection
with linear prediction and joint features. In MIREX (2007), onset detection
contest. (Cited on page 79).

Lerdahl, F. and Jackendoff, R. (1983). A Generative Theory of Tonal Music.

MIT Press, Cambridge. (Cited on page 33).

Leveau, P. (2007). A multipitch detection algorithm using a sparse

decomposition with instrument-specific harmonic atoms. In MIREX (2007),
multiple f0 estimation and tracking contest. (Cited on page 160).

Leveau, P., Vincent, E., Richard, G., and Daudet, L. (2008). Instrument-specific
harmonic atoms for mid-level music representation. IEEE Trans. on Audio,
Speech, and Language Processing, 16(1):116 – 128. (Cited on pages xi and 71).

202
 BIBLIOGRAPHY

Li, Y. and Wang, D. L. (2007). Pitch detection in polyphonic music

using instrument tone models. In Proc. IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), volume II, pages 481–484,
Honolulu, HI. (Cited on page 74).

Lidy, T., Grecu, A., Rauber, A., Pertusa, A., Ponce de León, P. J., and Iñesta,
J. M. (2009). A multi-feature multi-classifier ensemble approach for audio
music classification. In MIREX (2009), audio genre classification contest,
Kobe, Japan. (Cited on page 4).

Lidy, T., Rauber, A., Pertusa, A., and Iñesta, J. M. (2007). Improving
genre classification by combination of audio and symbolic descriptors using a
transcription system. In Proc. of the 8th International Conference on Music
Information Retrieval (ISMIR), pages 61–66, Vienna, Austria. (Cited on
pages 4, 52, 119, 159, and 168).

Lidy, T., Rauber, A., Pertusa, A., Ponce de León, P. J., and Iñesta, J. M. (2008).
Audio Music Classification Using A Combination Of Spectral, Timbral,
Rhythmic, Temporal And Symbolic Features. In MIREX (2008), audio genre
classification contest, Philadelphia, PA. (Cited on page 4).

Lloyd, L. S. (1970). Music and Sound. Ayer Publishing. (Cited on page 48).

Maestre, E. and Gómez, E. (2005). Automatic characterization of dynamics and

articulation of expressive monophonic recordings. In AES Convention 118.
(Cited on page 18).

Maher, R. C. (1989). An Approach for the Separation of Voices in Composite

Musical Signals. PhD thesis, University of Illinois, IL, USA. (Cited on page
4).

Maher, R. C. (1990). Evaluation of a method for separating digitized duet

signals. Journal of Audio Engineering Society, 38:956–979. (Cited on pages
58, 65, and 130).

Maher, R. C. and Beauchamp, J. W. (1994). Fundamental frequency estimation

of musical signals using a two-way mismatch procedure. Journal of the
Acoustical Society of America, 95:2254–2263. (Cited on pages xi, 58, and 59).

Mallat, S. (1999). A wavelet tour of signal processing. Academic press, second

edition. (Cited on page 13).

Mallat, S. and Zhang, Z. (1993). Matching pursuits with time-frequency

dictionaries. IEEE Trans. on Signal Processing, 41(12):3397–3415. (Cited
on page 70).

203
BIBLIOGRAPHY

Marolt, M. (2001). SONIC: transcription of polyphonic piano music with neural

networks. In Procs. of Workshop on Current Research Directions in Computer
Music. (Cited on page 101).

Marolt, M. (2002). Automatic Transcription of Piano Music with Neural

Networks. PhD thesis, University of Ljubljana, Slovenia. (Cited on page
4).

Marolt, M. (2004a). A connectionist approach to automatic transcription of

polyphonic piano music. IEEE Trans. on Multimedia, 6:439–449. (Cited on
pages xi, 67, 69, 101, and 102).

Marolt, M. (2004b). Networks of adaptive oscillators for partial tracking and

transcription of music recordings. Journal of New Music Research, 33(1):49–
59. (Cited on pages 67, 69, and 101).

Marolt, M., Kavcic, A., and Privosnik, M. (2002). Neural networks for note onset
detection in piano music. In Proc. International Computer Music Conference
(ICMC), Gothenburg, Sweden. (Cited on page 81).

Martin, K. (1996). A blackboard system for automatic transcription of simple

polyphonic music. Technical Report 385, MIT Media Lab. (Cited on pages
xii, 74, 75, and 77).

Masri, P. and Bateman, A. (1996). Improved modelling of attack transients

in music analysis-resynthesis. In Proc. of the International Computer Music
Conference (ICMC), pages 100–103, Hong Kong. (Cited on page 24).

Mcaulay, R. and Quatieri, T. (1986). Speech analysis/synthesis based on a

sinusoidal representation. IEEE Trans. on Acoustics, Speech, and Signal
Processing, 34(4):744–754. (Cited on page 23).

McKay, C. (2003). Using blackboard systems for polyphonic transcription: A

literature review. Course Paper, McGill University. (Cited on page 74).

Meddis, R. and Hewitt, M. J. (1991a). Virtual Pitch and Phase Sensitivity of a

Computer Model of the Auditory Periphery. I: Pitch Identification. Journal of
the Acoustical Society of America, 89:2866–2882. (Cited on pages 59 and 60).

Meddis, R. and Hewitt, M. J. (1991b). Virtual Pitch and Phase Sensitivity of

a Computer Model of the Auditory Periphery. II: Phase sensitivity. Journal
of the Acoustical Society of America, 89:2883–2894. (Cited on page 59).

Meng, X. L. and Rubin, D. B. (1993). Maximum likelihood estimation via the

ECM algorithm. Biometrika, 80(2):267–278. (Cited on page 74).

204
 BIBLIOGRAPHY

Meredith, D. and Wiggins, G. A. (2005). Comparing pitch spelling algorithms.

In International Symposium of Music Information Retrieval (ISMIR), pages
280–287, London, UK. (Cited on page 36).

Minsky, M. L. and Papert, S. (1969). Perceptrons. MIT Press, Cambridge, MA.

(Cited on page 38).

MIREX (2005). Music Information Retrieval Evaluation eXchange. Onset de-

tection contest. http://www.music-ir.org/evaluation/mirex-results/
audio-onset/index.html. (Cited on pages 93, 94, 192, 194, 202, and 208).

MIREX (2006). Music Information Retrieval Evaluation eXchange. Onset

detection contest. http://www.music-ir.org/mirex/2006/index.php/
Audio_Onset_Detection_Results. (Cited on page 93).

MIREX (2007). Music Information Retrieval Evaluation eXchange.

Multiple fundamental frequency estimation and tracking contest. http:
//www.music-ir.org/mirex/2007/index.php/Multiple_Fundamental_
Frequency_Estimation_%26_Tracking_Results. (Cited on pages xv, 52,
119, 127, 141, 157, 158, 159, 160, 161, 162, 167, 168, 193, 194, 197, 202, 207,
and 212).

MIREX (2007). Music Information Retrieval Evaluation eXchange. Onset

detection contest. http://www.music-ir.org/mirex/2007/index.php/
Audio_Onset_Detection_Results. (Cited on pages 93 and 202).

MIREX (2008). Music Information Retrieval Evaluation eXchange.

Multiple fundamental frequency estimation and tracking contest. http:
//www.music-ir.org/mirex/2008/index.php/Multiple_Fundamental_
Frequency_Estimation_%26_Tracking_Results. (Cited on pages xv, 52,
119, 141, 157, 161, 162, 163, 164, 167, 168, 193, 196, 197, 199, 207, 208, 213,
and 214).

MIREX (2009). Music Information Retrieval Evaluation eXchange. Onset

detection contest. http://www.music-ir.org/mirex/2009/index.php/
Audio_Onset_Detection_Results. (Cited on pages 93, 94, 95, 98, 207,
209, and 211).

Monti, G. and Sandler, M. B. (2002). Automatic polyphonic piano note

extraction using fuzzy logic in a blackboard system. In Proc. of the 5th
Conference on Digital Audio Effects (DAFx), Hamburg, Germany. (Cited on
page 74).

Moon, T. K. (1996). The Expectation-Maximization Algorithm. IEEE Signal

Processing Magazine, 13(6):47–60. (Cited on page 58).

205
BIBLIOGRAPHY

Moore, B. C. J., editor (1995). Hearing: Handbook of perception and cognition.

Academic Press, second edition. (Cited on page 17).

Moore, B. C. J. (1997). An introduction to the Psychology of Hearing. Academic

Press, fifth edition. (Cited on pages 10 and 53).

Moorer, J. A. (1975). On the Segmentation and Analysis of Continuous Musical

Sound by Digital Computer. PhD thesis, Dept. of Music, Stanford University.
(Cited on pages 4 and 62).

Moorer, J. A. (1977). Signal processing aspects of computer music: A survey.

Computer Music Journal, 1(1):4–37. (Cited on page 23).

Noll, M. (1967). Cepstrum pitch determination. Journal of the Acoustical

Society of America, 41:293–309. (Cited on page 56).

Noll, M. (1969). Pitch determination of human speech by the harmonic product

spectrum, the harmonic sum spectrum, and a maximum likelihood estimate.
In Proc. of the Symposium on Computer Processing Communications, pages
779–797, Polytechnic Institute of Brooklyn. (Cited on pages xi and 58).

Noll, M. and Schroeder, M. R. (1964). Short-time “cepstrum” pitch detection.

Journal of the Acoustical Society of America, 36:1030–1030. (Cited on page
56).

Paiement, J. C., Grandvalet, Y., and Bengio, S. (2008). Predictive models for
music. Connection Science, 21:253–272. (Cited on page 44).

Patterson, R. D. (1982). Auditory filter shapes derived with noise stimuli.

Journal of the Acoustical Society of America, 76:640–654. (Cited on page
17).

Patterson, R. D., Allerhand, M., and Giguere, C. (1995). Time-domain

modelling of peripheral auditory processing: A modular architecture and
software platform. Journal of the Acoustical Society of America, 98:1890–
1894. (Cited on page 17).

Patterson, R. D., Nimmo-Smith, I., Weber, D. L., and Milroy, R. (1982). The
deterioration of hearing with age: Frequency selectivity, the critical ratio,
the audiogram, and speech threshold. Journal of the Acoustical Society of
America, 72:1788–1803. (Cited on page 17).

Peeters, G. (2004). A large set of audio features for sound description (similarity
and classification) in the CUIDADO project. Technical report, IRCAM, Paris,
France. (Cited on pages 18, 19, and 20).

206
 BIBLIOGRAPHY

Peeters, G. (2006). Music pitch representation by periodicity measures

based on combined temporal and spectral representations. In Proc. of the
IEEE International Conference on Acoustics, Speech and Signal Processing
(ICASSP), volume V, pages 53–56, Toulouse, France. (Cited on page 63).

Pérez-Sancho, C. (2009). Stochastic Language Models for Music Information

Retrieval. PhD thesis, Universidad de Alicante, Spain. (Cited on page 169).

Pertusa, A. (2003). Transcripción de melodı́as polifónicas mediante redes

neuronales dinámicas. MSc Thesis, Universidad de Alicante, Spain. (Cited
on pages 101, 108, and 109).

Pertusa, A. and Iñesta, J. M. (2004). Pattern recognition algorithms for

polyphonic music transcription. In Fred, A., editor, Pattern Recognition in
Information Systems (PRIS), pages 80–89, Porto, Portugal. (Cited on page
101).

Pertusa, A. and Iñesta, J. M. (2005). Polyphonic monotimbral music

transcription using dynamic networks. Pattern Recognition Letters,
26(12):1809–1818. (Cited on page 101).

Pertusa, A. and Iñesta, J. M. (2007). Multiple fundamental frequency estimation

based on spectral pattern loudness and smoothness. In MIREX (2007),
multiple f0 estimation and tracking contest. (Cited on page 119).

Pertusa, A. and Iñesta, J. M. (2008a). Multiple fundamental frequency

estimation using Gaussian smoothness. In Proc. of the IEEE Int. Conf.
on Acoustics, Speech, and Signal Processing (ICASSP), pages 105–108, Las
Vegas, NV. (Cited on pages 119, 159, and 160).

Pertusa, A. and Iñesta, J. M. (2008b). Multiple fundamental frequency

estimation using Gaussian smoothness and short context. In MIREX (2008),
multiple f0 estimation and tracking contest. (Cited on pages 119, 163, 164,
and 167).

Pertusa, A. and Iñesta, J. M. (2009). Note onset detection using one semitone
filter-bank for MIREX 2009. In MIREX (2009), onset detection contest.
(Cited on pages 84, 95, and 96).

Pertusa, A., Klapuri, A., and Iñesta, J. M. (2005). Recognition of note onsets
in digital music using semitone bands. Lecture Notes in Computer Science,
3773:869–879. (Cited on pages 84 and 92).

Plumbley, M. D., Abdallah, S., Bello, J. P., Daview, M., Monti, G., and Sandler,
M. (2002). Automatic music transcription and audio source separation.
Cybernetic and Systems, 33(6):603–627. (Cited on pages 69, 70, and 74).

207
BIBLIOGRAPHY

Poliner, G. E. and Ellis, D. P. W. (2007a). A discriminative model for polyphonic

piano transcription. EURASIP Journal on Advances in Signal Processing.
(Cited on pages xi, 50, 51, 68, 157, 159, 160, 163, and 165).
Poliner, G. E. and Ellis, D. P. W. (2007b). Improving generalization for
classification-based polyphonic piano transcription. In Proc. of the IEEE
Workshop on Applications of Signal Processing to Audio and Acoustics
(WASPAA), pages 86–89, New Paltz, NY. (Cited on page 69).
Rabiner, L. (1989). A tutorial on hidden markov models and selected
applications in speech recognition. Proceedings of the IEEE, 77(2):257–286.
(Cited on page 61).
Rabiner, L., Cheng, M., Rosenberg, A., and McGonegal, C. (1976). A
comparative performance study of several pitch detection algorithms. IEEE
Trans. on Acoustics, Speech and Signal Processing, 24(5):399–418. (Cited on
pages 55 and 57).
Raczynski, S. A., Ono, N., and Sagayama, S. (2007). Multipitch analisys
with harmonic nonnegative matrix approximation. In Proc. of the 8th Int.
Conference on Music Information Retrieval (ISMIR), pages 381–386. (Cited
on pages 70 and 160).
Raphael, C. (2002). Automatic transcription of piano music. In Proc. Int.
Symposium on Music Information Retrieval (ISMIR), pages 15–19. (Cited
on page 72).
Reis, G., Fernandez, F., and Ferreira, A. (2008a). Genetic algorithm approach to
polyphonic music transcription for MIREX 2008. In MIREX (2008), multiple
f0 estimation and tracking contest. (Cited on pages 163 and 164).
Reis, G., Fonseca, N., de Vega, F., and Ferreira, A. (2008b). A genetic algorithm
based on gene fragment competition for polyphonic music transcription.
Lecture Notes in Computer Science, 4974:305–314. (Cited on page 68).
Reis, G., Fonseca, N., Fernandez, F., and Ferreira, A. (2008c). A genetic
algorithm approach with harmonic structure evolution for polyphonic music
transcription. IEEE International Symposium on Signal Processing and
Information Technology (ISSPIT), pages 491 – 496. (Cited on pages 67
and 68).
Roads, C. (1996). The Computer Music Tutorial. MIT Press, Cambridge.
(Cited on page 60).
Röbel, A. (2005). Onset detection in polyphonic signals by means of transient
peak classification. In MIREX (2005), onset detection contest. (Cited on
pages 80 and 94).

208
 BIBLIOGRAPHY

Röbel, A. (2009). Onset detection by means of transient peak classification in

harmonic bands. In MIREX (2009), onset detection contest. (Cited on pages
94, 95, 96, and 97).

Rodet, X. (1997). Musical sound signals analysis/synthesis: Sinusoidal+residual

and elementary waveform models. In Proc. of the IEEE Time-Frequency and
Time-Scale Workshop (TFTS’97), Coventry, GB. (Cited on page 121).

Rodet, X., Escribe, J., and Durignon, S. (2004). Improving score to audio
alignment: Percussion alignment and precise onset estimation. In Proc. of
Int. Computer Music Conference (ICMC), pages 450–453. (Cited on page
53).

Ruiz-Reyes, N., Vera-Candeas, P., nadas Quesada, F. J. C., and Carabias, J. J.

(2009). Fast communication: New algorithm based on spectral distance
maximization to deal with the overlapping partial problem in note-event
detection. Signal Processing, 89(8):1653–1660. (Cited on page 71).

Rumelhart, D. E., Hinton, G. E., and Williams, R. J. (1986). Learning

representations by back-propagating errors. Nature, 323:533–536. (Cited
on pages 39 and 104).

Ryynänen, M. (2006). Singing transcription. In Klapuri and Davy (2006),

chapter 12. (Cited on page 26).

Ryynänen, M. (2008). Automatic Transcription of Pitch Content in Music and

Selected Applications. PhD thesis, Tampere University of Technology. (Cited
on page 4).

Ryynänen, M. and Klapuri, A. (2004). Modelling of note events for singing

transcription. In in Proc. ISCA Tutorial and Research Workshop on Statistical
and Perceptual Audio, page 6. MIT Press. (Cited on pages xi, 56, 61, 64,
and 65).

Ryynänen, M. and Klapuri, A. (2005). Polyphonic music transcription using

note event modeling. In Proc. IEEE Workshop on Applications of Signal
Processing to Audio and Acoustics (WASPAA), pages 319–322, New Paltz,
New York, USA. (Cited on pages xi, 51, 64, 74, 120, 159, 160, 163, 164, 167,
and 169).

Sachs, C. (1940). The history of Musical Instruments. Norton, New York.

(Cited on page 20).

Sano, H. and Jenkins, B. K. (1989). A neural network model for pitch perception.
Computer Music Journal, 13(3):41–48. (Cited on page 102).

209
BIBLIOGRAPHY

Scheirer, E. D. (1998). Tempo and beat analysis of acoustic musical signals.

Journal of the Acoustical Society of America, 103(1):588–601. (Cited on
pages 83 and 85).

Schmidt, M. N. (2008). Single-channel source separation using non-negative

matrix factorization. PhD thesis, Technical University of Denmark. (Cited
on page 70).

Schouten, J. F. (1940). The residue and the mechanism of hearing.

In Proceedings Koninklijke Nederlandse Akademie van Wetenschappen,
volume 43, pages 991–999. (Cited on page 22).

Schwefel, H. P. (1995). Evolution and Optimum Seeking. Wiley & Sons, New
York. (Cited on page 66).

Selfbridge-Field, E. (1997). Beyond MIDI: the handbook of musical codes. MIT

Press, Cambridge, USA. (Cited on page 37).

Serra, X. (1997). Musical sound modeling with sinusoids plus noise. In Roads,
C., Pope, S. T., Picialli, A., and De Poli, G., editors, Musical signal processing,
pages 91–122. Swets and Zeitlinger. (Cited on pages 23 and 59).

Shannon, B. J. and Paliwal, K. K. (2003). A comparative study of filter bank

spacing for speech recognition. In Proc. Microelectronic Engineering Research
Conference. (Cited on page 16).

Slaney, M. (1993). An efficient implementation of the Patterson-Holdsworth

auditory filter bank. Technical Report 35, Perception Group, Advanced
Technology Group, Apple Computer, Inc. (Cited on page 17).

Sloboda, J. A. (1985). The musical mind. The cognitive psychology of music.

Oxford: The Clarendon Press. (Cited on page 43).

Smaragdis, P. (2001). Redundancy Reduction for Computational Audition, a

Unifying Approach. PhD thesis, MAS Department, MIT. (Cited on page
44).

Smaragdis, P. and Brown, J. (2003). Non-negative matrix factorization for

polyphonic music transcription. In Proc. IEEE Workshop on Applications
of Signal Processing to Audio and Acoustics (WASPAA), New Paltz, NY.
(Cited on pages xi, 69, and 70).

Sondhi, M. M. (1968). New methods of pitch extraction. IEEE Trans. Audio

Electroacoust., 16:262–266. (Cited on page 56).

210
 BIBLIOGRAPHY

Sterian, A. D. (1999). Model-Based Segmentation of Time-Frequency Images

for Musical Transcription. PhD thesis, University of Michigan. (Cited on
page 72).

Stevens, S., Volkman, J., and Newman, E. (1937). A scale for the measurement
of the psychological magnitude of pitch. Journal of the Acoustical Society of
America, 8(3):185–190. (Cited on page 16).

Stowell, D. and Plumbley, M. D. (2007). Adaptive whitening for improved real-

time audio onset detection. In Proc. of the Int. Computer Music Conference
(ICMC), pages 312–319. (Cited on pages 78 and 79).

Sundberg, J. (1987). The science of singing voice. Northern Illinois University

Press. (Cited on page 26).

Tan, H. L., Zhu, Y., and Chaisorn, L. (2009). An energy-based and pitch-based
approach to audio onset detection. In MIREX (2009), onset detection contest.
(Cited on pages 52, 82, 94, 95, 96, 97, and 99).

Taylor, I. and Greenhough, M. (1993). An object oriented ARTMAP system

for classifying pitch. In Proceedings of the International Computer Music
Conference (ICMC), pages 244–247, Tokyo, Japan. (Cited on page 102).

Tolonen, A. and Karjalainen, M. (2000). A computationally efficient multipitch

analysis model. IEEE Trans. on Speech and Audio Processing, 8(6):708–716.
(Cited on pages 60 and 62).

Tzanetakis, G. (2009). Marsyas submissions to MIREX 2009. In MIREX (2009),

onset detection contest. (Cited on pages 93, 95, and 96).

Tzanetakis, G., Essl, G., and Cook, P. (2001). Audio analysis using the discrete
wavelet transform. In Proc. Conf. in Acoustics and Music Theory Applications
(WSES). (Cited on page 13).

Vercoe, B. (1991). The CSound Reference Manual. MIT Press, Cambridge,

Massachusetts. (Cited on page 37).

Verma, T. S. and Meng, T. H. Y. (2000). Extending spectral modeling synthesis

with transient modeling synthesis. Computer Music Journal, 24(2):47–59.
(Cited on page 24).

Vidal, E., Casacuberta, F., Rodrı́guez, L., Civera, J., and Martı́nez, C. D.
(2006). Computer-assisted translation using speech recognition. IEEE Trans.
on Audio, Speech and Language Processing, 14(3):941–951. (Cited on page
173).

211
BIBLIOGRAPHY

Viitaniemi, T., Klapuri, A., and Eronen, A. (2003). A probabilistic model

for the transcription of single-voice melodies. In Proc. of the Finnish Signal
Processing Symposium (FINSIG), pages 59–63. (Cited on page 61).

Vincent, E. (2004). Modèles d’instruments pour la séparation de sources et la

transcription d’enregistrements musicaux. PhD thesis, Université Paris VI.
(Cited on page 4).

Vincent, E., Bertin, N., and Badeau, R. (2007). Two nonnegative matrix
factorization methods for polyphonic pitch transcription. In MIREX (2007),
multiple f0 estimation and tracking contest. (Cited on pages 70, 159, 160,
163, and 164).

Vincent, E. and Plumbley, M. D. (2005). A prototype system for object coding of

musical audio. In Proc. IEEE Workshop on Applications of Signal Processing
to Audio and Acoustics (WASPAA), pages 249–242, New Paltz, NY. (Cited
on page 72).

Vincent, E. and Rodet, X. (2004). Music transcription with ISA and HMM. In
Proc. 5th International Conference on Independent Component Analysis and
Blind Signal Separation, pages 1197–1204. (Cited on page 72).

Virtanen, T. (2000). Audio signal modeling with sinusoids plus noise. MSc
Thesis, Tampere University of Technology. (Cited on pages 121 and 122).

Virtanen, T. (2006). Unsupervised learning methods for source separation. In

Klapuri and Davy (2006), chapter 9. (Cited on page 70).

Virtanen, T. (2007). Monaural sound source separation by nonnegative matrix

factorization with temporal continuity and sparseness criteria. IEEE Trans.
on Audio, Speech and Language Processing, 15(3):1066–1074. (Cited on page
70).

Viterbi, A. (1967). Error bounds for convolutional codes and an asymptotically

optimum decoding. IEEE Trans. on Information Theory, 13(2):260–269.
(Cited on page 61).

Vos, P. and Vianen, B. G. V. (1984). Thresholds for discrimination between

pure and tempered intervals: The relevance of nearly coinciding harmonics.
Journal of the Acoustical Society of America, 77:176–187. (Cited on page
30).

Waibel, A. (1989). Modular construction of time-delay neural networks for

speech recognition. Neural Computation, 1:39–46. (Cited on pages 39
and 102).

212
 BIBLIOGRAPHY

Walmsley, P., Godsill, S., and Rayner, P. (1999). Polyphonic pitch tracking
using joint bayesian estimation of multiple frame parameters. In Proc.
IEEE Workshop on Applications of Signal Processing to Audio and Acoustics
(WASPAA), pages 119–122, New Paltz, NY. (Cited on page 72).

Wan, J., Wu, Y., and Dai, H. (2005). A harmonic enhancement based multipitch
estimation algorithm. In IEEE International Symposium on Communications
and Information Technology (ISCIT) 2005, volume 1, pages 772 – 776. (Cited
on page 65).

Wang, W., Luo, Y., Chambers, J. A., and Sanei, S. (2008). Note onset detection
via nonnegative factorization of magnitude spectrum. EURASIP Journal on
Advances in Signal Processing. (Cited on page 81).

Wessel, D. L. (1979). Timbre space as a musical control structure. Computer

Music Journal, 3(2):45–52. (Cited on page 19).

Wood, A. (2008). The physics of music. Davies Press. (Cited on page 47).

Woodruff, J., Li, Y., and Wang, D. (2008). Resolving overlapping harmonics
fpr monoaural musical sound separation using pitch and common amplitude
modulation. In Proc. of the International Symposium on Music Information
Retrieval (ISMIR), pages 538–543, Philadelphia, PA. (Cited on page 130).

Yeh, C. (2008). Multiple fundamental frequency estimation of polyphonic

recordings. PhD thesis, Universite Paris VI - Pierre et Marie Curie. (Cited
on pages xi, 4, 45, 46, 48, 65, 76, 128, 160, and 173).

Yeh, C., Röbel, A., and Rodet, X. (2005). Multiple fundamental frequency
estimation of polyphonic music signals. In IEEE, Int. Conf. on Acoustics,
Speech and Signal Processing (ICASSP), volume III, pages 225–228,
Philadelphia, PA. (Cited on pages 65, 126, 127, 129, and 130).

Yeh, C. and Roebel, A. (2006). Adaptive noise level estimation. In Proc. of

the 9th Int. Conference on Digital Audio Effects (DAFx), Montreal, Canada.
(Cited on page 65).

Yeh, C. and Roebel, A. (2009). The expected amplitude of overlapping partials

of harmonic sounds. In Proc. of the International Conference on Acoustics,
Speech and Signal Processing (ICASSP), Taipei, Taiwan. (Cited on page 47).

Yeh, C., Roebel, A., and Chang, W. C. (2008). Multiple F0 estimation for
MIREX 08. In MIREX (2008), multiple f0 estimation and tracking contest.
(Cited on pages 163, 164, and 167).

213
BIBLIOGRAPHY

Yeh, C., Roebel, A., and Rodet, X. (2006). Multiple f0 tracking in solo
recordings of monodic instruments. In Proc. of the 120th AES Convention,
Paris, France. (Cited on pages 48 and 66).

Yin, J., Sim, T., Wang, Y., and Shenoy, A. (2005). Music transcription using
an instrument model. In Proc. of the IEEE International Conference on
Acoustics, Speech, and Signal Processing (ICASSP), volume III, pages 217–
220. (Cited on page 65).

Young, S., Kershaw, D., Odell, J., Ollason, D., Valtchev, V., and Woodland, P.
(2000). The HTK book (for HTK version 3.1). Cambridge University. (Cited
on page 87).

Young, S. J., Russell, N. H., and Thornton, J. H. S. (1989). Token passing: a

simple conceptual model for connected speech recognition systems. Technical
report, Cambridge University Engineering Department. (Cited on page 65).

Zhou, R. (2006). Feature extraction of Musical Content For Automatic Music

Transcription. PhD thesis, École Polytechnique Féderale de Laussane. (Cited
on pages 4 and 69).

Zhou, R. and Mattavelli, M. (2007). A new time-frequency representation

for music signal analysis: Resonator time-frequency image. In Proc. Int.
Conference on Information Sciences, Signal Processing and its Applications,
Sharijah, U. Arab Emirates. (Cited on pages 63 and 80).

Zhou, R., Mattavelli, M., and Zoia, G. (2008). Music Onset Detection Based
on Resonator Time Frequency Image. IEEE Transactions On Audio, Speech
And Language Processing, 16(8):1685–1695. (Cited on page 80).

Zhou, R. and Reiss, J. D. (2008). A real-time polyphonic music transcription

system. In MIREX (2008), multiple f0 estimation and tracking contest.
(Cited on page 164).

Zhou, R., Reiss, J. D., Mattavelli, M., and Zoia, G. (2009). A computationally
efficient method for polyphonic pitch estimation. EURASIP Journal on
Advances in Signal Processing, (28). (Cited on pages 63, 127, and 160).

Zhu, Y. and Kankanhalli, M. (2006). Precise pitch profile feature extraction from
musical audio for key detection. IEEE Trans. on Multimedia, 8(3):575–584.
(Cited on pages 94 and 99).

Zwicker, E. (1961). Subdivision of the audible frequency range into critical bands
(Frequenzgruppen). Journal of the Acoustical Society of America, 33(2):248.
(Cited on page 16).

214
 BIBLIOGRAPHY

Zwicker, E., Flottorp, G., and Stevens, S. S. (1957). Critical bandwidth in

loudness summation. Journal of the Acoustical Society of America, 29:548–
557. (Cited on page 16).

215