P1: IML/FFX P2: IML

MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

Dynamic Speech Models

Theory, Algorithms, and Applications

i
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

Copyright © 2006 by Morgan & Claypool

All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in
any form or by any means—electronic, mechanical, photocopy, recording, or any other except for brief quotations
in printed reviews, without the prior permission of the publisher.

Dynamic Speech Models, Theory, Algorithms, and Applications

Li Deng
www.morganclaypool.com

1598290649 paper Deng

1598290657 ebook Deng

DOI: 10.2200/S00028ED1V01Y200605SAP002

A Publication in the Morgan & Claypool Publishers’ series

SYNTHESIS LECTURES ON SPEECH AND AUDIO PROCESSING
Lecture #2
Series editor B. H. Juang

First Edition
10 9 8 7 6 5 4 3 2 1

Printed in the United States of America

ii
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

Dynamic Speech Models

Theory, Algorithms, and Applications
Li Deng
Microsoft Research
Redmond, Washington, USA

SYNTHESIS LECTURES ON SPEECH AND AUDIO PROCESSING #2

M
&C Mor gan & Cl aypool Publishers

iii
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

iv

ABSTRACT
Speech dynamics refer to the temporal characteristics in all stages of the human speech com-
munication process. This speech “chain” starts with the formation of a linguistic message in a
speaker’s brain and ends with the arrival of the message in a listener’s brain. Given the intri-
cacy of the dynamic speech process and its fundamental importance in human communication,
this monograph is intended to provide a comprehensive material on mathematical models of
speech dynamics and to address the following issues: How do we make sense of the complex
speech process in terms of its functional role of speech communication? How do we quantify
the special role of speech timing? How do the dynamics relate to the variability of speech that
has often been said to seriously hamper automatic speech recognition? How do we put the
dynamic process of speech into a quantitative form to enable detailed analyses? And finally,
how can we incorporate the knowledge of speech dynamics into computerized speech analysis
and recognition algorithms? The answers to all these questions require building and applying
computational models for the dynamic speech process.
What are the compelling reasons for carrying out dynamic speech modeling? We pro-
vide the answer in two related aspects. First, scientific inquiry into the human speech code has
been relentlessly pursued for several decades. As an essential carrier of human intelligence and
knowledge, speech is the most natural form of human communication. Embedded in the speech
code are linguistic (as well as para-linguistic) messages, which are conveyed through four levels
of the speech chain. Underlying the robust encoding and transmission of the linguistic mes-
sages are the speech dynamics at all the four levels. Mathematical modeling of speech dynamics
provides an effective tool in the scientific methods of studying the speech chain. Such scientific
studies help understand why humans speak as they do and how humans exploit redundancy and
variability by way of multitiered dynamic processes to enhance the efficiency and effectiveness
of human speech communication. Second, advancement of human language technology, espe-
cially that in automatic recognition of natural-style human speech is also expected to benefit
from comprehensive computational modeling of speech dynamics. The limitations of current
speech recognition technology are serious and are well known. A commonly acknowledged and
frequently discussed weakness of the statistical model underlying current speech recognition
technology is the lack of adequate dynamic modeling schemes to provide correlation structure
across the temporal speech observation sequence. Unfortunately, due to a variety of reasons,
the majority of current research activities in this area favor only incremental modifications and
improvements to the existing HMM-based state-of-the-art. For example, while the dynamic
and correlation modeling is known to be an important topic, most of the systems neverthe-
less employ only an ultra-weak form of speech dynamics; e.g., differential or delta parameters.
Strong-form dynamic speech modeling, which is the focus of this monograph, may serve as an
ultimate solution to this problem.
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

After the introduction chapter, the main body of this monograph consists of four chapters.
They cover various aspects of theory, algorithms, and applications of dynamic speech models,
and provide a comprehensive survey of the research work in this area spanning over past 20 years.
This monograph is intended as advanced materials of speech and signal processing for graudate-
level teaching, for professionals and engineering practioners, as well as for seasoned researchers
and engineers specialized in speech processing.

KEYWORDS
Articulatory trajectories, Automatic speech recognition, Coarticulation, Discretizing hidden
dynamics, Dynamic Bayesian network, Formant tracking, Generative modeling, Speech
acoustics, Speech dynamics, Vocal tract resonance

v
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

vi
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

vii

Contents
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 What Are Speech Dynamics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 What Are Models of Speech Dynamics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Why Modeling Speech Dynamics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 Outline of the Book . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2. A General Modeling and Computational Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2.1 Background and Literature Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Model Design Philosophy and Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Model Components and the Computational Framework . . . . . . . . . . . . . . . . . . . . 13
2.3.1 Overlapping Model for Multitiered Phonological Construct . . . . . . . . . 13
2.3.2 Segmental Target Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.3.3 Articulatory Dynamic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.4 Functional Nonlinear Model for Articulatory-to-
Acoustic Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.5 Weakly Nonlinear Model for Acoustic Distortion . . . . . . . . . . . . . . . . . . . 24
2.3.6 Piecewise Linearized Approximation for Articulatory-to-
Acoustic Mapping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3. Modeling: From Acoustic Dynamics to Hidden Dynamics . . . . . . . . . . . . . . . . . . . . . . 31

3.1 Background and Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2 Statistical Models for Acoustic Speech Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2.1 Nonstationary-State HMMs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.2 Multiregion Recursive Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Statistical Models for Hidden Speech Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3.1 Multiregion Nonlinear Dynamic System Models . . . . . . . . . . . . . . . . . . . . 36
3.3.2 Hidden Trajectory Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4. Models with Discrete-Valued Hidden Speech Dynamics . . . . . . . . . . . . . . . . . . . . . . . . 39

4.1 Basic Model with Discretized Hidden Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.1.1 Probabilistic Formulation of the Basic Model . . . . . . . . . . . . . . . . . . . . . . . 40
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

viii CONTENTS
4.1.2 Parameter Estimation for the Basic Model: Overview . . . . . . . . . . . . . . . 41
4.1.3 EM Algorithm: The E-Step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.1.4 A Generalized Forward-Backward Algorithm . . . . . . . . . . . . . . . . . . . . . . 43
4.1.5 EM Algorithm: The M-Step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.6 Decoding of Discrete States by Dynamic Programming. . . . . . . . . . . . . .48
4.2 Extension of the Basic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2.1 Extension from First-Order to Second-Order Dynamics . . . . . . . . . . . . . 49
4.2.2 Extension from Linear to Nonlinear Mapping . . . . . . . . . . . . . . . . . . . . . . 50
4.2.3 An Analytical Form of the Nonlinear Mapping Function . . . . . . . . . . . . 51
4.2.4 E-Step for Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4.2.5 M-Step for Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2.6 Decoding of Discrete States by Dynamic Programming. . . . . . . . . . . . . .61
4.3 Application to Automatic Tracking of Hidden Dynamics . . . . . . . . . . . . . . . . . . . 61
4.3.1 Computation Efficiency: Exploiting Decomposability in the
Observation Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.3.2 Experimental results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

5. Models with Continuous-Valued Hidden Speech Trajectories . . . . . . . . . . . . . . . . . . . 69

5.1 Overview of the Hidden Trajectory Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.1.1 Generating Stochastic Hidden Vocal Tract
Resonance Trajectories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.1.2 Generating Acoustic Observation Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.1.3 Linearizing Cepstral Prediction Function . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.1.4 Computing Acoustic Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2 Understanding Model Behavior by Computer Simulation . . . . . . . . . . . . . . . . . . . 76
5.2.1 Effects of Stiffness Parameter on Reduction . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2.2 Effects of Speaking Rate on Reduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . .78
5.2.3 Comparisons with Formant Measurement Data . . . . . . . . . . . . . . . . . . . . . 79
5.2.4 Model Prediction of Vocal Tract Resonance Trajectories
for Real Speech Utterances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.2.5 Simulation Results on Model Prediction for Cepstral
Trajectories. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .82
5.3 Parameter Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.3.1 Cepstral Residuals’ Distributional Parameters . . . . . . . . . . . . . . . . . . . . . . . 84
5.3.2 Vocal Tract Resonance Targets’ Distributional Parameters . . . . . . . . . . . 89
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

CONTENTS ix
5.4 Application to Phonetic Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.4.1 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.4.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

x
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

xi

Acknowledgments
This book would not have been possible without the help and support from friends, family,
colleagues, and students. Some of the material in this book is the result of collaborations with
my former students and current colleagues. Special thanks go to Jeff Ma, Leo Lee, Dong Yu,
Alex Acero, Jian-Lai Zhou, and Frank Seide.
The most important acknowledgments go to my family. I also thank Microsoft Research
for providing the environment in which the research described in this book is made possible.
Finally, I thank Prof. Fred Juang and Joel Claypool for not only the initiation but also the
encouragement and help throughout the course of writting this book.
P1: IML/FFX P2: IML
MOBK024-FM MOBK024-LiDeng.cls May 24, 2006 8:16

xii
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK024-01 MOBK024-LiDeng.cls May 24, 2006 8:6

CHAPTER 1

Introduction

1.1 WHAT ARE SPEECH DYNAMICS?

In a broad sense, speech dynamics are time-varying or temporal characteristics in all stages
of the human speech communication process. This process, sometimes referred to as speech
chain [1], starts with the formation of a linguistic message in the speaker’s brain and ends with
the arrival of the message in the listener’s brain. In parallel with this direct information transfer,
there is also a feedback link from the acoustic signal of speech to the speaker’s ear and brain.
In the conversational mode of speech communication, the style of the speaker’s speech can be
further influenced by an assessment of the extent to which the linguistic message is successfully
transferred to or understood by the listener. This type of feedbacks makes the speech chain a
closed-loop process.
The complexity of the speech communication process outlined above makes it desirable to
divide the entire process into modular stages or levels for scientific studies. A common division of
the direct information transfer stages of the speech process, which this book is mainly concerned
with, is as follows:

• Linguistic level: At this highest level of speech communication, the speaker forms the
linguistic concept or message to be conveyed to the listener. That is, the speaker decides
to say something linguistically meaningful. This process takes place in the language
center(s) of speaker’s brain. The basic form of the linguistic message is words, which are
organized into sentences according to syntactic constraints. Words are in turn composed
of syllables constructed from phonemes or segments, which are further composed of
phonological features. At this linguistic level, language is represented in a discrete or
symbolic form.
• Physiological level: Motor program and articulatory muscle movement are involved at
this level of speech generation. The speech motor program takes the instructions, spec-
ified by the segments and features formed at the linguistic level, on how the speech
sounds are to be produced by the articulatory muscle (i.e., articulators) movement
over time. Physiologically, the motor program executes itself by issuing time-varying
commands imparting continuous motion to the articulators including the lips, tongue,
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK024-01 MOBK024-LiDeng.cls May 24, 2006 8:6

2 DYNAMIC SPEECH MODELS

larynx, jaw, and velum, etc. This process involves coordination among various articu-
lators with different limitations in the movement speed, and it also involves constant
corrective feedback. The central scientific issue at this level is how the transformation
is accomplished from the discrete linguistic representation to the continuous articula-
tors’ movement or dynamics. This is sometimes referred to as the problem of interface
between phonology and phonetics.
• Acoustic level: As a result of the articulators’ movements, acoustic air stream emerges
from the lungs, and passes through the vocal cords where a phonation type is developed.
The time-varying sound sources created in this way are then filtered by the time-varying
acoustic cavities shaped by the moving articulators in the vocal tract. The dynamics of
this filter can be mathematically represented and approximated by the changing vocal
tract area function over time for many practical purposes. The speech information
at the acoustic level is in the form of dynamic sound pattern after this filtering process.
The sound wave radiated from the lips (and in some cases from the nose and through the
tissues of the face) is the most accessible element of the multiple-level speech process
for practical applications. For example, this speech sound wave may be easily picked
by a microphone and be converted to analog or digital electronic form for storage or
transmission. The electronic form of speech sounds makes it possible to transport them
thousands of miles away without loss of fidelity. And computerized speech recognizers
gain access to speech data also primarily in the electronic form of the original acoustic
sound wave.
• Auditory and perceptual level: During human speech communication, the speech sound
generated at the acoustic level above impinges upon the eardrums of a listener, where it
is first converted to mechanical motion via the ossicles of the middle ear, then to fluid
pressure waves in the medium bathing the basilar membrane of the inner ear invoking
traveling waves. This finally excites hair cells’ electrical, mechanical, and biochemical
activities, causing firings in some 30,000 human auditory nerve fibers. These various
stages of the processing carry out some nonlinear form of frequency analysis, with the
analysis results in the form of dynamic spatial–temporal neural response patterns. The
dynamic spatial–temporal neural responses are then sent to higher processing centers
in the brain, including the brainstem centers, the thalamus, and the primary auditory
cortex. The speech representation in the primary auditory cortex (with a high degree
of plasticity) appears to be in the form of multiscale and jointly spectro-temporally
modulated patterns. For the listener to extract the linguistic content of speech, a process
that we call speech perception or decoding, it is necessary to identify the segments and
features that underlie the sound pattern based on the speech representation in the
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK024-01 MOBK024-LiDeng.cls May 24, 2006 8:6

INTRODUCTION 3
primary auditory cortex. The decoding process may be aided by some type of analysis-
by-synthesis strategies that make use of general knowledge of the dynamic processes at
the physiological and acoustic levels of the speech chain as the “encoder” device for the
intended linguistic message.
At all the four levels of the speech communication process above, dynamics play a central
role in shaping the linguistic information transfer. At the linguistic level, the dynamics are
discrete and symbolic, as is the phonological representation. That is, the discrete phonological
symbols (segments or features) change their identities at various points of time in a speech
utterance, and no quantitative (numeric) degree of change and precise timing are observed.
This can be considered as a weak form of dynamics. In contrast, the articulatory dynamics at
the physiological level, and the consequent dynamics at the acoustic level, are of a strong form
in that the numerically quantifiable temporal characteristics of the articulator movements and
of the acoustic parameters are essential for the trade-off between overcoming the physiological
limitations for setting the articulators’ movement speed and efficient encoding of the phono-
logical symbols. At the auditory level, the importance of timing in the auditory nerve’s firing
patterns and in the cortical responses in coding speech has been well known. The dynamic
patterns in the aggregate auditory neural responses to speech sounds in many ways reflect the
dynamic patterns in the speech signal, e.g., time-varying spectral prominences in the speech
signal. Further, numerous types of auditory neurons are equipped with special mechanisms (e.g.,
adaptation and onset-response properties) to enhance the dynamics and information contrast
in the acoustic signal. These properties are especially useful for detecting certain special speech
events and for identifying temporal “landmarks” as a prerequisite for estimating the phonological
features relevant to consonants [2, 3].
Often, we use our intuition to appreciate speech dynamics—as we speak, we sense the
motions of speech articulators and the sounds generated from these motions as continuous flow.
When we call this continuous flow of speech organs and sounds as speech dynamics, then we
use them in a narrow sense, ignoring their linguistic and perceptual aspects.
As is often said, timing is of essence in speech. The dynamic patterns associated with ar-
ticulation, vocal tract shaping, sound acoustics, and auditory response have the key property that
the timing axis in these patterns is adaptively plastic. That is, the timing plasticity is flexible but
not arbitrary. Compression of time in certain portions of speech has a significant effect in speech
perception, but not so for other portions of the speech. Some compression of time, together
with the manipulation of the local or global dynamic pattern, can change perception of the style
of speaking but not the phonetic content. Other types of manipulation, on the other hand, may
cause very different effects. In speech perception, certain speech events, such as labial stop bursts,
flash extremely quickly over as short as 1–3 ms while providing significant cues for the listener
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK024-01 MOBK024-LiDeng.cls May 24, 2006 8:6

4 DYNAMIC SPEECH MODELS

to identify the relevant phonological features. In contrast, for other phonological features, even
dropping a much longer chunk of the speech sound would not affect their identification. All
these point to the very special status of time in speech dynamics. The time in speech seems to
be quite different from the linear flow of time as we normally experience it in our living world.
Within the speech recognition community, researchers often refer to speech dynamics
as differential or regression parameters derived from the acoustic vector sequence (called delta,
delta–delta, or “dynamic” features) [4, 5]. From the perspective of the four-level speech chain
outlined above, such parameters can at best be considered as an ultra-weak form of speech
dynamics. We call them ultra-weak not only because they are confined to the acoustic domain
(which is only one of the several stages in the complete speech chain), but also because temporal
differentiation can be regarded hardly as a full characterization in the actual dynamics even
within the acoustic domain. As illustrated in [2, 6, 7], the acoustic dynamics of speech exhib-
ited in spectrograms have the intricate, linguistically correlated patterns far beyond what the
simplistic differentiation or regression can characterize. Interestingly, there have been numer-
ous publications on how the use of the differential parameters is problematic and inconsistent
within the traditional pattern recognition frameworks and how one can empirically remedy
the inconsistency (e.g., [8]). The approach that we will describe in this book gives the subject
of dynamic speech modeling a much more comprehensive and rigorous treatment from both
scientific and technological perspectives.

1.2 WHAT ARE MODELS OF SPEECH DYNAMICS?

As discussed above, the speech chain is a highly dynamic process, relying on the coordination
of linguistic, articulatory, acoustic, and perceptual mechanisms that are individually dynamic as
well. How do we make sense of this complex process in terms of its functional role of speech
communication? How do we quantify the special role of speech timing? How do the dynamics
relate to the variability of speech that has often been said to seriously hamper automatic speech
recognition? How do we put the dynamic process of speech into a quantitative form to enable
detailed analyses? How can we incorporate the knowledge of speech dynamics into computerized
speech analysis and recognition algorithms? The answers to all these questions require building
and applying computational models for the dynamic speech process.
A computational model is a form of mathematical abstraction of the realistic physical
process. It is frequently established with necessary simplification and approximation aimed at
mathematical or computational tractability. The tractability is crucial in making the mathemat-
ical abstraction amenable to computer or algorithmic implementation for practical engineering
applications. Applying this principle, we define models of speech dynamics in the context of this
book as the mathematical characterization and abstraction of the physical speech dynamics.
These characterization and abstraction are capable of capturing the essence of time-varying
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK024-01 MOBK024-LiDeng.cls May 24, 2006 8:6

INTRODUCTION 5
aspects in the speech chain and are sufficiently simplified to facilitate algorithm development
and engineering system implementation for speech processing applications. It is highly desirable
that the models be developed in statistical terms, so that advanced algorithms can be developed
to automatically and optimally determine any parameters in the models from a representative
set of training data. Further, it is important that the probability for each speech utterance be
efficiently computed under any hypothesized word-sequence transcript to make the speech
decoding algorithm development feasible.
Motivated by the multiple-stage view of the dynamic speech process outlined in the
preceding section, detailed computational models, especially those for the multiple generative
stages, can be constructed from the distinctive feature-based linguistic units to acoustic and
auditory parameters of speech. These stages include the following:

• A discrete feature-organization process that is closely related to speech gesture over-

lapping and represents partial or full phone deletion and modifications occurring per-
vasively in casual speech;
• a segmental target process that directs the model-articulators up-and-down and front-
and-back movements in a continuous fashion;
• the target-guided dynamics of model-articulators movements that flow smoothly from
one phonological unit to the next; and
• the static nonlinear transformation from the model-articulators to the measured speech
acoustics and the related auditory speech representations.

The main advantage of modeling such detailed multiple-stage structure in the dynamic
human speech process is that a highly compact set of parameters can then be used to cap-
ture phonetic context and speaking rate/style variations in a unified framework. Using this
framework, many important subjects in speech science (such as acoustic/auditory correlates of
distinctive features, articulatory targets/dynamics, acoustic invariance, and phonetic reduction)
and those in speech technology (such as modeling pronunciation variation, long-span context-
dependence representation, and speaking rate/style modeling for recognizer design) that were
previously studied separately by different communities of researchers can now be investigated
in a unified fashion.
Many aspects of the above multitiered dynamic speech model class, together with its scien-
tific background, have been discussed in [9]. In particular, the feature organization/overlapping
process, as is central to a version of computational phonology, has been presented in some
detail under the heading of “computational phonology.” Also, some aspects of auditory speech
representation, limited mainly to the peripheral auditory system’s functionalities, have been
elaborated in [9] under the heading of “auditory speech processing.” This book will treat these
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK024-01 MOBK024-LiDeng.cls May 24, 2006 8:6

6 DYNAMIC SPEECH MODELS

topics only lightly, especially considering that both computational phonology and high-level
auditory processing of speech are still active ongoing research areas. Instead, this book will
concentrate on the following:

• The target-based dynamic modeling that interfaces between phonology and

articulation-based phonetics;
• the switching dynamic system modeling that represents the continuous, target-directed
movement in the “hidden” articulators and in the vocal tract resonances being closely
related to the articulatory structure; and
• the relationship between the “hidden” articulatory or vocal tract resonance parame-
ters to the measurable acoustic parameters, enabling the hidden speech dynamics to
be mapped stochastically to the acoustic dynamics that are directly accessible to any
machine processor.

In this book, these three major components of dynamic speech modeling will be treated
in a much greater depth than in [9], especially in model implementation and in algorithm
development. In addition, this book will include comprehensive reviews of new research work
since the publication of [9] in 2003.

1.3 WHY MODELING SPEECH DYNAMICS?

What are the compelling reasons for carrying out dynamic speech modeling? We provide
the answer in two related aspects. First, scientific inquiry into the human speech code has
been relentlessly pursued for several decades. As an essential carrier of human intelligence
and knowledge, speech is the most natural form of human communication. Embedded in the
speech code are linguistic (and para-linguistic) messages, which are conveyed through the four
levels of the speech chain outlined earlier. Underlying the robust encoding and transmission of
the linguistic messages are the speech dynamics at all the four levels (in either a strong form
or a weak form). Mathematical modeling of the speech dynamics provides one effective tool
in the scientific methods of studying the speech chain—observing phenomena, formulating
hypotheses, testing the hypotheses, predicting new phenomena, and forming new theories.
Such scientific studies help understand why humans speak as they do and how humans exploit
redundancy and variability by way of multitiered dynamic processes to enhance the efficiency
and effectiveness of human speech communication.
Second, advancement of human language technology, especially that in automatic recog-
nition of natural-style human speech (e.g., spontaneous and conversational speech), is also
expected to benefit from comprehensive computational modeling of speech dynamics. Auto-
matic speech recognition is a key enabling technology in our modern information society. It
serves human–computer interaction in the most natural and universal way, and it also aids the
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK024-01 MOBK024-LiDeng.cls May 24, 2006 8:6

INTRODUCTION 7
enhancement of human–human interaction in numerous ways. However, the limitations of
current speech recognition technology are serious and well known (e.g., [10–13]). A commonly
acknowledged and frequently discussed weakness of the statistical model (hidden Markov model
or HMM) underlying current speech recognition technology is the lack of adequate dynamic
modeling schemes to provide correlation structure across the temporal speech observation se-
quence [9, 13, 14]. Unfortunately, due to a variety of reasons, the majority of current research
activities in this area favor only incremental modifications and improvements to the existing
HMM-based state-of-the-art. For example, while the dynamic and correlation modeling is
known to be an important topic, most of the systems nevertheless employ only the ultra-weak
form of speech dynamics, i.e., differential or delta parameters. A strong form of dynamic speech
modeling presented in this book appears to be an ultimate solution to the problem.
It has been broadly hypothesized that new computational paradigms beyond the conven-
tional HMM as a generative framework are needed to reach the goal of all-purpose recognition
technology for unconstrained natural-style speech, and that statistical methods capitalizing
on essential properties of speech structure are beneficial in establishing such paradigms. Over
the past decade or so, there has been a popular discriminant-function-based and conditional
modeling approach to speech recognition, making use of HMMs (as a discriminant function
instead of as a generative model) or otherwise [13, 15–19]. This approach has been grounded
on the assumption that we do not have adequate knowledge about the realistic speech process,
as exemplified by the following quote from [17]: “The reason of taking a discriminant function
based approach to classifier design is due mainly to the fact that we lack complete knowledge
of the form of the data distribution and training data are inadequate.” The special difficulty of
acquiring such distributional speech knowledge lies in the sequential nature of the data with a
variable and high dimensionality. This is essentially the problem of dynamics in the speech data.
As we gradually fill in such knowledge while pursing research in dynamic speech modeling, we
will be able to bridge the gap between the discriminative paradigm and the generative modeling
one, but with a much higher performance level than the systems at present. This dynamic speech
modeling approach can enable us to “put speech science back into speech recognition” instead of
treating speech recognition as a generic, loosely constrained pattern recognition problem. In this
way, we are able to develop models “that really model speech,” and such models can be expected to
provide an opportunity to lay a foundation of the next-generation speech recognition technology.

1.4 OUTLINE OF THE BOOK

After the introduction chapter, the main body of this book consists of four chapters. They cover
theory, algorithms, and applications of dynamic speech models and survey in a comprehensive
manner the research work in this area spanning over past 20 years or so. In Chapter 2, a general
framework for modeling and for computation is presented. It provides the design philosophy
for dynamic speech models and outlines five major model components, including phonological
P1: IML/FFX P2: IML/FFX QC: IML/FFX T1: IML
MOBK024-01 MOBK024-LiDeng.cls May 24, 2006 8:6

8 DYNAMIC SPEECH MODELS

construct, articulatory targets, articulatory dynamics, acoustic dynamics, and acoustic distor-
tions. For each of these components, relevant speech science literatures are discussed, and general
mathematical descriptions are developed with needed approximations introduced and justified.
Dynamic Bayesian networks are exploited to provide a consistent probabilistic language for
quantifying the statistical relationships among all the random variables in the dynamic speech
models, including both within-component and cross-component relationships.
Chapter 3 is devoted to a comprehensive survey of many different types of statistical mod-
els for speech dynamics, from the simple ones that focus on only the observed acoustic patterns
to the more advanced ones that represent the dynamics internal to the surface acoustic domain
and represent the relationship between these “hidden” dynamics and the observed acoustic dy-
namics. This survey classifies the existing models into two main categories—acoustic dynamic
models and hidden dynamic models, and provides a unified perspective viewing these models as
having different degrees of approximation to the realistic multicomponent overall speech chain.
Within each of these two main model categories, further classification is made depending
on whether the dynamics are mathematically defined with or without temporal recursion.
Consequences of this difference in the algorithm development are addressed and discussed.
Chapters 4 and 5 present two types of hidden dynamic models that are best developed
to date as reported in the literature, with distinct model classes and distinct approximation and
implementation strategies. They exemplify the state-of-the-arts in the research area of dynamic
speech modeling. The model described in Chapter 4 uses discretization of the hidden dynamic
variables to overcome the original difficulty of intractability in algorithms for parameter estima-
tion and for decoding the phonological states. Modeling accuracy is inherently limited to the
discretization precision, and the new computation difficulty arising from the large discretiza-
tion levels due to multi-dimensionality in the hidden dynamic variables is addressed by a greedy
optimization technique. Except for these two approximations, the parameter estimation and
decoding algorithms developed and described in this chapter are based on rigorous EM and
dynamic programming techniques. Applications of this model and the related algorithms to
the problem of automatic hidden vocal tract resonance tracking are presented, where the esti-
mates are for the discretized hidden resonance values determined by the dynamic programming
technique for decoding based on the EM-trained model parameters.
The dynamic speech model presented in Chapter 5 maintains the continuous nature in the
hidden dynamic values, and uses an explicit temporal function (i.e., defined nonrecursively) to
represent the hidden dynamics or “trajectories.” The approximation introduced to overcome the
original intractability problem is made by iteratively refining the boundaries associated with
the discrete phonological units while keeping the boundaries fixed when carrying out parameter
estimation. We show computer simulation results that demonstrate the desirable model behavior
in characterizing coarticulation and phonetic reduction. Applications to phonetic recognition
are also presented and analyzed.
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

CHAPTER 2

A General Modeling and

Computational Framework

The main aim of this chapter is to set up a general modeling and computational framework,
based on the modern mathematical tool called dynamic Bayesian networks (DBN) [20, 21],
and to establish general forms of the multistage dynamic speech model outlined in the preceding
chapter. The overall model presented within this framework is comprehensive in nature, covering
all major components in the speech generation chain—from the multitiered, articulation-based
phonological construct (top) to the environmentally distorted acoustic observation (bottom).
The model is formulated in specially structured DBN, in which the speech dynamics at separate
levels of the generative chain are represented by distinct (but dependent) discrete and continuous
state variables and by their characteristic temporal evolution.
Before we present the model and the associated computational framework, we first provide
a general background and literature review.

2.1 BACKGROUND AND LITERATURE REVIEW

In recent years, the research community in automatic speech recognition has started attacking
the difficult problem in the research field—conversational and spontaneous speech recogni-
tion (e.g., [12, 16, 22–26]). This new endeavor has been built upon the dramatic progress in
speech technology achieved over the past two decades or so [10, 27–31]. While the progress
has already created many state-of-the-art speech products, ranging from free-text dictation
to wireless access of information via voice, the conditions under which the technology works
well are highly limited. The next revolution in speech recognition technology should enable
machines to communicate with humans in a natural and unconstrained way. To achieve this
challenging goal, some of researchers (e.g., [3, 10, 11, 13, 20, 22, 32–39]) believe that the se-
vere limitations of the HMM should be overcome and novel approaches to representing key
aspects of the human speech process are highly desirable or necessary. These aspects, many of
which are of a dynamic nature, have been largely missing in the conventional hidden Markov
model (HMM) based framework. Towards this objective, one specific strategy would be to
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

10 DYNAMIC SPEECH MODELS

place appropriate dynamic structure on the speech model that allows for the kinds of variations
observed in conversational speech. Furthermore, enhanced computational methods, including
learning and inference techniques, will also be needed based on new or extended models beyond
the HMM.
It has been well known that many assumptions in the current HMM framework are
inconsistent with the underlying mechanisms, as well as with the surface acoustic observations,
of the dynamic speech process (e.g., [13, 14]). However, a number of approaches, notably
known as the stochastic segment model, segmental HMM, and trended trajectory models
[14,40,41], which are intended to specifically overcome some of these inconsistent assumptions,
have not delivered significant and consistent performance improvements, especially in large
speech recognition systems. Most of these approaches aimed mainly at overcoming the HMM’s
assumption of conditional IID (independent and identical distribution conditioned on the
HMM state sequence). Yet the inconsistency between the HMM assumptions and the properties
of the realistic dynamic speech process goes far beyond this “local” IID limitation. It appears
necessary not just to empirically fix one weak aspect of the HMM or another, but rather to
develop the new computational machinery that directly incorporates key dynamic properties of
the human speech process. One natural (but challenging) way of incorporating such knowledge
is to build comprehensive statistical generative models for the speech dynamics. Via the use of
Bayes theorem, the statistical generative models explicitly give the posterior probabilities for
different speech classes, enabling effective decision rules to be constructed in a speech recognizer.
In discriminative modeling approaches, on the other hand, where implicit computation of the
posterior probabilities for speech classes is carried out, it is generally much more difficult to
systematically incorporate knowledge of the speech dynamics.
Along the direction of generative modeling, many researchers have, over recent years, been
proposing and pursuing research that extensively explores the dynamic properties of speech in
various forms and at various levels of the human speech communication chain (e.g., [24,32–34,
42–49]). Some approaches have advocated the use of the multitiered feature-based phonological
units, which control human speech production and are typical of human lexical representation
(e.g., [11, 50–52]). Other approaches have emphasized the functional significance of abstract,
“task” dynamics in speech production and recognition (e.g., [53, 54]). Yet other approaches
have focused on the dynamic aspects in the speech process, where the dynamic object being
modeled is in the space of surface speech acoustics, rather than in the space of the intermediate,
production-affiliated variables that are internal to the direct acoustic observation (e.g., [14, 26,
55–57]).
Although dynamic modeling has been a central focus of much recent work in speech
recognition, the dynamic object being modeled, either in the space of “task” variables or of
acoustic variables, does not, and potentially may not be able to, directly take into account
the many important properties in realistic articulatory dynamics. Some earlier proposals and
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 11

empirical methods for modeling pseudo-articulatory dynamics or abstract hidden dynamics
for the purpose of speech recognition can be found in [42, 44–46, 58, 59]. In these studies,
the dynamics of a set of pseudo-articulators are realized either by filtering from sequentially
arranged, phoneme-specific target positions or by applying trajectory-smoothness constraints.
Due to the simplistic nature in the use of the pseudo-articulators, one important property
of human speech production, compensatory articulation, could not be taken into account,
because it would require modeling correlations among target positions of a set of articu-
lators. This has reduced the power of such models for potentially successful use in speech
recognition.
To incorporate essential properties in human articulatory dynamics—including compen-
satory articulation, target-directed behavior, and flexibly constrained dynamics due to biome-
chanical properties of different articulatory organs—in a statistical generative model of speech, it
appears necessary to use essential properties of realistic multidimensional articulators. Previous
attempts using the pseudo-articulators did not incorporate most of such essential properties. Be-
cause much of the acoustic variation observed in speech that makes speech recognition difficult
can be attributed to articulatory phenomena, and because articulation is one key component in
the closed-loop human speech communication chain, it is highly desirable to develop an explicit
articulation-motivated dynamic model and to incorporate it into a comprehensive generative
model of the dynamic speech process.
The comprehensive generative model of speech and the associated computational frame-
work discussed in this chapter consists of a number of key components that are centered on
articulatory dynamics. A general overview of this multicomponent model is provided next, fol-
lowed by details of the individual model components including their mathematical descriptions
and their DBN representations.

2.2 MODEL DESIGN PHILOSOPHY AND OVERVIEW

Spontaneous speech (e.g., natural voice mails and lectures) and speech of verbal conversations
among two or more speakers (e.g., over the telephone or in meetings) are pervasive forms of
human communication. If a computer system can be constructed to automatically decode the
linguistic messages contained in spontaneous and conversational speech, one will have vast
opportunities for the applications of speech technology.
What characterizes spontaneous and conversational speech is its casual style nature. The
casual style of speaking produces two key consequences that make the acoustics of spontaneous
and conversational speech significantly differ from that of the “read-style” speech: phono-
logical reorganization and phonetic reduction. First, in casual speech, which is called “hypo-
articulated” speech in [60], phonological reorganization occurs where the relative timing or
phasing relationship across different articulatory feature/gesture dimensions in the “orches-
trated” feature strands are modified. One obvious manifestation of this modification is that the
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

12 DYNAMIC SPEECH MODELS

more casual or relaxed the speech style is, the greater the overlapping across the feature/gesture
dimensions becomes. Second, phonetic reduction occurs where articulatory targets as phonetic
correlates to the phonological units may shift towards a more neutral position due to the use of
reduced articulatory efforts. Phonetic reduction also manifests itself by pulling the realized ar-
ticulatory trajectories further away from reaching their respective targets due to physical inertia
constraints in the articulatory movements. This occurs within generally shorter time duration
in casual-style speech than in the read-style speech.
It seems difficult for the HMM systems to provide effective mechanisms to embrace
the huge, new acoustic variability in casual, spontaneous, and conversational speech arising
either from phonological organization or from phonetic reduction. Importantly, the additional
variability due to phonetic reduction is scaled continuously, resulting in phonetic confusions
in a predictable manner. (See Chapter 5 for some detailed computation simulation results
pertaining to such prediction.) Due to this continuous variability scaling, very large amounts
of (labeled) speech data would be needed. Even so, they can only partly capture the variability
when no structured knowledge about phonetic reduction and about its effects on speech dynamic
patterns is incorporated into the speech model underlying spontaneous and conversational
speech-recognition systems.
The general design philosophy of the mathematical model for the speech dynamics de-
scribed in this chapter is based on the desire to integrate the structured knowledge of both
phonological reorganization and phonetic reduction. To fully describe this model, we break up
the model into several interrelated components, where the output, expressed as the probability
distribution, of one component serves as the input to the next component in a “generative” spirit.
That is, we characterize each model component as a joint probability distribution of both input
and output sequences, where both the sequences may be hidden. The top-level component is the
phonological model that specifies the discrete (symbolic) pronunciation units of the intended
linguistic message in terms of multitiered, overlapping articulatory features. The first intermedi-
ate component consists of articulatory control and target, which provides the interface between
the discrete phonological units to the continuous phonetic variable and which represents the
“ideal” articulation and its inherent variability if there were no physical constraints in articu-
lation. The second intermediate component consists of articulatory dynamics, which explicitly
represents the physical constraints in articulation and gives the output of “actual” trajectories in
the articulatory variables. The bottom component would be the process of speech acoustics be-
ing generated from the vocal tract whose shape and excitation are determined by the articulatory
variables as the output of the articulatory dynamic model. However, considering that such a
clean signal is often subject to one form of acoustic distortion or another before being processed
by a speech recognizer, and further that the articulatory behavior and the subsequent speech
dynamics in acoustics may be subject to change when the acoustic distortion becomes severe,
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 13

we complete the comprehensive model by adding the final component of acoustic distortion
with feedback to the higher level component describing articulatory dynamics.

2.3 MODEL COMPONENTS AND THE COMPUTATIONAL

FRAMEWORK
In a concrete form, the generative model for speech dynamics, whose design philosophy and
motivations have been outlined in the preceding section, consists of the hierarchically structured
components of
1. multitiered phonological construct (nonobservable or hidden; discrete valued);
2. articulatory targets (hidden; continuous-valued);
3. articulatory dynamics (hidden; continuous);
4. acoustic pattern formation from articulatory variables (hidden; continuous); and
5. distorted acoustic pattern formation (observed; continuous).

In this section, we will describe each of these components and their design in some detail.
In particular, as a general computational framework, we provide the DBN representation for
each of the above model components and for their combination.

2.3.1 Overlapping Model for Multitiered Phonological Construct

Phonology is concerned with sound patterns of speech and with the nature of the discrete
or symbolic units that shapes such patterns. Traditional theories of phonology differ in the
choice and interpretation of the phonological units. Early distinctive feature-based theory [61]
and subsequent autosegmental, feature-geometry theory [62] assumed a rather direct link be-
tween phonological features and their phonetic correlates in the articulatory or acoustic domain.
Phonological rules for modifying features represented changes not only in the linguistic struc-
ture of the speech utterance, but also in the phonetic realization of this structure. This weakness
has been recognized by more recent theories, e.g., articulatory phonology [63], which empha-
size the importance of accounting for phonetic levels of variation as distinct from those at the
phonological levels.
In the phonological model component described here, it is assumed that the linguis-
tic function of phonological units is to maintain linguistic contrasts and is separate from
phonetic implementation. It is further assumed that the phonological unit sequence can be
described mathematically by a discrete-time, discrete-state, multidimensional homogeneous
Markov chain. How to construct sequences of symbolic phonological units for any arbitrary
speech utterance and how to build them into an appropriate Markov state (i.e., phonological
state) structure are two key issues in the model specification. Some earlier work on effective
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

14 DYNAMIC SPEECH MODELS

methods of constructing such overlapping units, either by rules or by automatic learning, can
be found in [50, 59, 64–66]. In limited experiments, these methods have proved effective for
coarticulation modeling in the HMM-like speech recognition framework (e.g., [50, 65]).
Motivated by articulatory phonology [63], the asynchronous, feature-based phonological
model discussed here uses multitiered articulatory features/gestures that are temporally over-
lapping with each other in separate tiers, with learnable relative-phasing relationships. This
contrasts with most existing speech-recognition systems where the representation is based on
phone-sized units with one single tier for the phonological sequence acting as “beads-on-a-
string.” This contrast has been discussed in some detail in [11] with useful insight.
Mathematically, the L-tiered, overlapping model can be described by the “factorial”
Markov chain [51, 67], where the state of the chain is represented by a collection of discrete-
component state variables for each time frame t:

st = s t(1) , . . . , s t(l) , . . . , s t(L) .

Each of the component states can take K (l) values. In implementing this model for American
English, we have L = 5, and the five tiers are Lips, Tongue Blade, Tongue Body, Velum, and
Larynx, respectively. For “Lips” tier, we have K (1) = 6 for six possible linguistically distinct Lips
configurations, i.e., those for /b/, /r/, /sh/, /u/, /w/, and /o/. Note that at this phonological level,
the difference among these Lips configurations is purely symbolic. The numerical difference is
manifested in different articulatory target values at lower phonetic level, resulting ultimately in
different acoustic realizations. For the remaining tiers, we have K (2) = 6, K (3) = 17, K (4) = 2,
and K (5) = 2.
The state–space of this factorial Markov chain consists of all K L = K (1) × K (2) × K (3) ×
K (4) × K (5) possible combinations of the s t(l) state variables. If no constraints are imposed on
the state transition structure, this would be equivalent to the conventional one-tiered Markov
chain with a total of K L states and a K L × K L state transition matrix. This would be an unin-
teresting case since the model complexity is exponentially (or factorially) growing in L. It would
also be unlikely to find any useful phonological structure in this huge Markov chain. Further,
since all the phonetic parameters in the lower level components of the overall model (to be
discussed shortly) are conditioned on the phonological state, the total number of model param-
eters would be unreasonably large, presenting a well-known sparseness difficulty for parameter
learning.
Fortunately, rich sources of phonological and phonetic knowledge are available to
constrain the state transitions of the above factorial Markov chain. One particularly useful set
of constraints come directly from the phonological theories that motivated the construction
of this model. Both autosegmental phonology [62] and articulatory phonology [63] treat the
different tiers in the phonological features as being largely independent of each other in their
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 15

S1(1) S 2(1) S3(1) S4(1) ....... ST(1)

S1( 2) S 2( 2) S 3( 2) S4(2 ) ....... S T(2)

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

S1(L ) S2( L ) S3(L ) S4( L ) ....... S T( L )

FIGURE 2.1: A dynamic Bayesian network (DBN) for a constrained factorial Markov chain as a prob-
abilistic model for an L-tiered overlapping phonological model based on articulatory features/gestures.
The constrained transition structure in the factorial Markov chain makes different tiers in the phono-
logical features independent of each other in their evolving dynamics. This gives rise to parallel streams,
s (l) , l = 1, 2, . . . , L, of the phonological features in their associated articulatory dimensions

evolving dynamics. This thus allows the a priori decoupling among the L tiers:


L
P (st |st−1 ) = P (s t(l) |s t−1
(l)
).
l=1

The transition structure of this constrained (uncoupling) factorial Markov chain can be
parameterized by L distinct K (l) × K (l) matrices. This is significantly simpler than the original
K L × K L matrix as in the unconstrained case.
Fig. 2.1 shows a dynamic Bayesian network (DBN) for a factorial Markov chain with
the constrained transition structure. A Bayesian network is a graphical model that describes
dependencies and conditional independencies in the probabilistic distributions defined over a
set of random variables. The most interesting class of Bayesian networks, as relevant to speech
modeling, is the DBN specifically aimed at modeling time series data or symbols such as speech
acoustics, phonological units, or a combination of them. For the speech data or symbols, there
are causal dependencies between random variables in time and they are naturally suited for the
DBN representation.
In the DBN representation of Fig. 2.1 for the L-tiered phonological model, each node
represents a component phonological feature in each tier as a discrete random variable at a
particular discrete time. The fact that there is no dependency (lacking arrows) between the
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

16 DYNAMIC SPEECH MODELS

nodes in different tiers indicates that each tier is autonomous in the evolving dynamics. We
call this model the overlapping model, reflecting the independent dynamics of the features
at different tiers. The dynamics cause many possible combinations in which different feature
values associated with their respective tiers occur simultaneously at a fixed time point. These are
determined by how the component features/gestures overlap with each other as a consequence
of their independent temporal dynamics. Contrary to this view, in the conventional phone-
based phonological model, there is only one single tier of phones as the “bundled” component
features, and hence there is no concept of overlapping component features.
In a DBN, the dependency relationships among the random variables can be implemented
by specifying the associated conditional probability for each node given all its parents. Because
of the decoupled structure across the tiers as shown in Fig. 2.1, the horizontal (temporal)
dependency is the only dependency that exists for the component phonological (discrete) states.
This dependency can be specified by the Markov chain transition probability for each separate
tier, l, defined by
 
P s t(l) = j |s t−1
(l)
= i = a i(l)j . (2.1)

2.3.2 Segmental Target Model

After a phonological model is constructed, the process for converting abstract phonological units
into their phonetic realization needs to be specified. The key issue here is whether the invariance
in the speech process is more naturally expressed in the articulatory or the acoustic/auditory
domain. A number of theories assumed a direct link between abstract phonological units and
physical measurements. For example, the “quantal theory” [68] proposed that phonological
features possessed invariant acoustic (or auditory) correlates that could be measured directly
from the speech signal. The “motor theory” [69] instead proposed that articulatory properties
are directly associated with phonological symbols. No conclusive evidence supporting either
hypothesis has been found without controversy [70].
In the generative model of speech dynamics discussed here, one commonly held view
in phonetics literature is adopted. That is, discrete phonological units are associated with a
temporal segmental sequence of phonetic targets or goals [71–75]. In this view, the function
of the articulatory motor control system is to achieve such targets or goals by manipulating the
articulatory organs according to some control principles subject to the articulatory inertia and
possibly minimal-energy constraints [60].
Compensatory articulation has been widely documented in the phonetics literature where
trade-offs between different articulators and nonuniqueness in the articulatory–acoustic map-
ping allow for the possibility that many different articulatory target configurations may be
able to “equivalently” realize the same underlying goal. Speakers typically choose a range
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 17

of possible targets depending on external environments and their interactions with listen-
ers [60, 70, 72, 76, 77]. To account for compensatory articulation, a complex phonetic control
strategy need to be adopted. The key modeling assumptions adopted here regarding such a strat-
egy are as follows. First, each phonological unit is correlated to a number of phonetic parameters.
These measurable parameters may be acoustic, articulatory, or auditory in nature, and they can
be computed from some physical models for the articulatory and auditory systems. Second, the
region determined by the phonetic correlates for each phonological unit can be mapped onto
an articulatory parameter space. Hence, the target distribution in the articulatory space can
be determined simply by stating what the phonetic correlates (formants, articulatory positions,
auditory responses, etc.) are for each of the phonological units (many examples are provided
in [2]), and by running simulations in detailed articulatory and auditory models. This particular
proposal for using the joint articulatory, acoustic, and auditory properties to specify the artic-
ulatory control in the domain of articulatory parameters was originally proposed in [59, 78].
Compared with the traditional modeling strategy for controlling articulatory dynamics [79]
where the sole articulatory goal is involved, this new strategy appears more appealing not only
because of the incorporation of the perceptual and acoustic elements in the specification of the
speech production goal, but also because of its natural introduction of statistical distributions
at the relatively high level of speech production.
A convenient mathematical representation for the distribution of the articulatory target
vector t follows a multivariate Gaussian distribution, denoted by

t ∼ N (t; m(s), (s)), (2.2)

where m(s) is the mean vector associated with the composite phonological state s, and
the covariance matrix (s) is nondiagonal. This allows for the correlation among the ar-
ticulatory vector components. Because such a correlation is represented for the articula-
tory target (as a random vector), compensatory articulation is naturally incorporated in the
model.
Since the target distribution, as specified in Eq. (2.2), is conditioned on a specific phono-
logical unit (e.g., a bundle of overlapped features represented by the composite state s consisting
of component feature values in the factorial Markov chain of Fig. 2.1), and since the target does
not switch until the phonological unit changes, the statistics for the temporal sequence of the
target process follows that of a segmental HMM [40].
For the single-tiered (L = 1) phonological model (e.g., phone-based model), the
segmental HMM for the target process will be the same as that described in [40], except
the output is no longer the acoustic parameters. The dependency structure in this segmental
HMM as the combined one-tiered phonological model and articulatory target model can be
illustrated in the DBN of Fig. 2.2. We now elaborate on the dependencies in Fig. 2.2. The
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

18 DYNAMIC SPEECH MODELS

S1 S2 S3 S4 SK

t1 t2 t3 t4 tK

FIGURE 2.2: DBN for a segmental HMM as a probabilistic model for the combined one-tiered phono-
logical model and articulatory target model. The output of the segmental HMM is the target vector, t,
constrained to be constant until the discrete phonological state, s , changes its value

output of this segmental HMM is the random articulatory target vector t(k) that is constrained
to be constant until the phonological state switches its value. This segmental constraint for
the dynamics of the random target vector t(k) represents the adopted articulatory control
strategy that the goal of the motor system is to try to maintain the articulatory target’s position
(for a fixed corresponding phonological state) by exerting appropriate muscle forces. That is,
although random, t(k) remains fixed until the phonological state s k switches. The switching of
target t(k) is synchronous with that of the phonological state, and only at the time of switching,
is t(k) allowed to take a new value according to its probability density function. This segmental
constraint can be described mathematically by the following conditional probability density
function:

δ[t(k) − t(k − 1)] if s k = s k−1 ,
p[t(k)|s k , s k−1 , t(k − 1)] =
N (t(k); m(s k ), (s k )) otherwise.

This adds the new dependencies of random vector of t(k) on s k−1 and t(k − 1), in addition to
the obvious direct dependency on s k , as shown in Fig. 2.2.
Generalizing from the one-tiered phonological model to the multitiered one as discussed
earlier, the dependency structure in the “segmental factorial HMM” as the combined multitiered
phonological model and articulatory target model has the DBN representation in Fig. 2.3. The
key conditional probability density function (PDF) is similar to the above segmental HMM
except that the conditioning phonological states are the composite states (sk and sk−1 ) consisting
of a collection of discrete component state variables:

δ[t(k) − t(k − 1)] if sk = sk−1 ,
p[t(k)|sk , sk−1 , t(k − 1)] =
N (t(k); m(sk ), (sk )) otherwise.

Note that in Figs. 2.2 and 2.3 the target vector t(k) is defined in the same space as that of
the physical articulator vector (including jaw positions, which do not have direct phonological
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 19

S1(1) S2(1) S 3(1) S4(1) ....... ST(1)

S1( 2 ) S2( 2) S3( 2 ) S4( 2) ....... ST

(2 )

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

S1( L ) S2( L ) S3( L ) S4( L ) ....... ST( L)

t1 t2 t3 t4 tK

FIGURE 2.3: DBN for a segmental factorial HMM as a combined multitiered phonological model and
articulatory target model

connections). And compensatory articulation can be represented directly by the articulatory

target distributions with a nondiagonal covariance matrix for component correlation. This
correlation shows how the various articulators can be jointly manipulated in a coordinated
manner to produce the same phonetically implemented phonological unit.
An alternative model for the segmental target model, as proposed in [33] and called the
“target dynamic” model, uses vocal tract constrictions (degrees and locations) instead of artic-
ulatory parameters as the target vector, and uses a geometrically defined nonlinear relationship
(e.g., [80]) to map one vector of vocal tract constrictions into a region (with a probability dis-
tribution) of the physical articulatory variables. In this case, compensatory articulation can also
be represented by the distributional region of the articulatory vectors induced indirectly by any
fixed vocal tract constriction vector.
The segmental factorial HMM presented here is a generalization of the segmental HMM
proposed originally in [40]. It is also a generalization of the factorial HMM that has been
developed from the machine learning community [67] and been applied to speech recognition
[51]. These generalizations are necessary because the output of our component model (not the
full model) is physically the time-varying articulatory targets as random sequences, rather than
the random acoustic sequences as in the segmental HMM or the factorial HMM.
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

20 DYNAMIC SPEECH MODELS

2.3.3 Articulatory Dynamic Model
Due to the difficulty in knowing how the conversion of higher-level motor control into artic-
ulator movement takes place, a simplifying assumption is made for the new model component
discussed in this subsection. That is, we assume that at the functional level the combined (non-
linear) control system and articulatory mechanism behave as a linear dynamic system. This
combined system attempts to track the control input, equivalently represented by the articu-
latory target, in the physical articulatory parameter space. Articulatory dynamics can then be
approximated as the response of a linear filter driven by a random target sequence as represented
by a segmental factorial HMM just described. The statistics of the random target sequence ap-
proximate those of the muscle forces that physically drives motions of the articulators. (The
output of this hidden articulatory dynamic model then produces a time-varying vocal tract
shape that modulates the acoustic properties of the speech signal.)
The above simplifying assumption then reduces the generally intractable nonlinear state
equation,

z(k + 1) = gs [z(k), ts , w(k)],

into the following mathematically tractable, linear, first-order autoregressive (AR) model:

z(k + 1) = As z(k) + Bs ts + w(k), (2.3)

where z is the n-dimensional real-valued articulatory-parameter vector, w is the IID and Gaus-
sian noise, ts is the HMM-state-dependent target vector expressed in the same articulatory
domain as z(k), As is the HMM-state-dependent system matrix, and Bs is a matrix that mod-
ifies the target vector. The dependence of ts and s parameters of the above dynamic system
on the phonological state is justified by the fact that the functional behavior of an articulator
depends both on the particular goal it is trying to implement, and on the other articulators with
which it is cooperating in order to produce compensatory articulation.
In order for the modeled articulatory dynamics above to exhibit realistic behaviors, e.g.,
movement along the target-directed path within each segment and not oscillating within the
segment, matrices As and Bs can be constrained appropriately. One form of the constraint gives
rise to the following articulatory dynamic model:

z(k + 1) = s z(k) + (I − s )ts + w(k), (2.4)

where I is the identity matrix. Other forms of the constraint will be discussed in Chapters 4
and 5 of the book for two specific implementions of the general model.
It is easy to see that the constrained linear AR model of Eq. (2.4) has the desirable target-
directed property. That is, the articulatory vector z(k) asymptotically approaches the mean of
the target random vector t for artificially lengthened speech utterances. For natural speech, and
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 21

especially for conversational speech with a casual style, the generally short duration associated
with each phonological state forces the articulatory dynamics to move away from the target
of the current state (and towards the target of the following phonological state) long before it
reaches the current target. This gives rise to phonetic reduction, and is one key source of speech
variability that is difficult to be directly captured by a conventional HMM.
Including the linear dynamic system model of Eq. (2.4), the combined phonological,
target, and articulatory dynamic model now has the DBN representation of Fig. 2.4. The new
dependency for the continuous articulatory state is specified, on the basis of Eq. (2.4), by the
following conditional PDF:

p z [z(k + 1)|z(k), t(k), sk ] = p w [z(k + 1) − sk z(k) − (I − sk )t(k)]. (2.5)

This combined model is a switching, target-directed AR model driven by a segmental factorial

HMM.

S1(1) S2(1) S 3(1) S4(1) ....... ST(1)

S1( 2 ) S2( 2) S3(2 ) S4( 2) ....... ST(2 )

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

S1( L ) S2( L ) S 3( L ) S4( L ) ....... ST( L)

t1 t2 t3 t4 tK

z1 z2 z3 z4 zK

FIGURE 2.4: DBN for a switching, target-directed AR model driven by a segmental factorial HMM.
This is a combined model for multitiered phonology, target process, and articulatory dynamics
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

22 DYNAMIC SPEECH MODELS

2.3.4 Functional Nonlinear Model for Articulatory-to-Acoustic Mapping
The next component in the overall model of speech dynamics moves the speech generative
process from articulation down to distortion-free speech acoustics. While a truly consistent
framework to accomplish this based on explicit knowledge of speech production should include
detailed mechanisms for articulatory-to-acoustic generation, this becomes impractical due to
difficulties in modeling learning and excessive computational requirements. Again, we make
the simplifying assumptions that the articulatory and acoustic states of the vocal tract can be
adequately described by low-order vectors of variables, where the articulatory state variables
represent respectively the relative positions of the major articulators, and the acoustic state
variables represent corresponding time-averaged spectral-like parameters computed from the
acoustic signal. If an appropriate time scale is chosen, the relationship between the articulatory
and acoustic representations can be modeled by a static memoryless transformation, converting
a vector of articulatory parameters into a vector of acoustic ones. This assumption appears
reasonable for a vocal tract with about 10-ms reverberation time.
This static memoryless transformation can be mathematically represented by the follow-
ing “observation” equation in the state–space model:
o(k) = h[z(k)] + v(k), (2.6)

where o is the m-dimensional real-valued observation vector, v is the IID observation noise
vector uncorrelated with the state noise w, and h[·] is the static memoryless transformation
from the articulatory vector to its corresponding acoustic observation vector.
Including this static mapping model, the combined phonological, target, articulatory
dynamic, and the acoustic model now has the DBN representation shown in Fig. 2.5. The new
dependency for the acoustic random variables is specified, on the basis of “observation” equation
in Eq. (2.6), by the following conditional PDF:
p o [o(k) | z(k)] = p v [o(k) − h(z(k))]. (2.7)
There are many ways of choosing the static nonlinear function for h[z] in Eq. (2.6), such as
using a multilayer perceptron (MLP) neural network. Typically, the analytical forms of nonlinear
functions make the associated nonlinear dynamic systems difficult to analyze and make the
estimation problems difficult to solve. Simplification is frequently used to gain computational
advantages while sacrificing accuracy for approximating the nonlinear functions. One most
commonly used technique for the approximation is truncated (vector) Taylor series expansion.
If all the Taylor series terms of order two and higher are truncated, then we have the linear
Taylor series approximation that is characterized by the Jacobian matrix J and by the point of
Taylor series expansion z0 :
h(z) ≈ h(z0 ) + J(z0 )(z − z0 ). (2.8)
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 23

S1(1) S2(1) S3(1) S4(1) ....... ST(1)

S1( 2 ) S2( 2) S3( 2 ) S4( 2) ....... ST(2 )

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

S1( L ) S2( L ) S3( L ) S4( L ) ....... ST( L)

t1 t2 t3 t4 tK

z1 z2 z3 z4 zK

o1 o2 o3 o4 oK

FIGURE 2.5: DBN for a target-directed, switching dynamic system (state–space) model driven by a seg-
mental factorial HMM. This is a combined model for multitiered phonology, target process, articulatory
dynamics, and articulatory-to-acoustic mapping

Each element of the Jacobian matrix J is partial derivative of each vector component of the
nonlinear output with respect to each of the input vector components. That is,

⎡ ⎤
∂h 1 (z0 ) ∂h 1 (z0 )
· · · ∂h∂z
1 (z0 )

⎢ ∂z1 ∂z2 n
⎥
∂h ⎢ ∂h 2 (z0 ) ∂h 2 (z0 )
· · · ∂h∂z
2 (z0 ) ⎥
⎢ ∂z1 ∂z2 ⎥
J(z0 ) = =⎢ n
⎥. (2.9)
∂z0 ⎢ .. .. ⎥
⎣ . . ⎦
∂h m (z0 ) ∂h m (z0 )
∂z1 ∂z2
· · · ∂h∂z
m (z0 )
n
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

24 DYNAMIC SPEECH MODELS

The radial basis function (RBF) is an attractive alternative to the MLP as a universal
function approximator [81] for implementing the articulatory-to-acoustic mapping. Use of the
RBF for a nonlinear function in the general nonlinear dynamic system model can be found
in [82], and will not be elaborated here.

2.3.5 Weakly Nonlinear Model for Acoustic Distortion

In practice, the acoustic vector of speech, o(k), generated from the articulatory vector z(k) is
subject to acoustic environmental distortion before being “observed” and being processed by a
speech recognizer for linguistic decoding. In many cases of interest, acoustic distortion can be
accurately characterized by joint additive noise and channel (convolutional) distortion in the
signal-sample domain. In this domain, the distortion is linear and can be described by

y(t) = o (t) ∗ (t) + n(t), (2.10)

where y(t) is the distorted speech signal sample, modeled as convolution between the clean
speech signal sample o (t) and distortion channel’s impulse response (t) plus additive noise
sample n(t).
However, in the log-spectrum domain or in the cepstrum domain that is commonly used
as the input for speech recognizers, Eq. (2.10) has its equivalent form of (see a derivation in [83])

y(k) = o(k) + h̄ + C log I + exp[C−1 (n(k) − o(k) − h̄)] . (2.11)

In Eq. (2.11), y(k) is the cepstral vector at frame k, when C is taken as cosine transform
matrix. (When C is taken as the identity matrix, y(k) becomes a log-spectral vector.) y(k) now
becomes weakly nonlinearly related to the clean speech cepstral vector o(k), cepstral vector of
additive noise n(k), and cepstral vector of the impulse response of the distortion channel h̄. Note
that according to Eq. (2.11), the relationship between clean and noisy speech cepstral vectors
becomes linear (or affine) when the signal-to-noise ratio is either very large or very small.
After incorporating the above acoustic distortion model, and assuming that the statistics of
the additive noise changes slowly over time as governed by a discrete-state Markov chain, Fig. 2.6
shows the DBN for the comprehensive generative model of speech from the phonological
model to distorted speech acoustics. Intermediate models include the target model, articulatory
dynamic model, and clean-speech acoustic model. For clarity, only a one-tiered, rather than
multitiered, phonological model is illustrated. [The dependency of the parameters () of the
articulatory dynamic model on the phonological state is also explicitly added.] Note that in Fig.
2.6, the temporal dependency in the discrete noise states Nk gives rise to nonstationarity in the
additive noise random vectors nk . The cepstral vector h̄ for the distortion channel is assumed
not to change over the time span of the observed distorted speech utterance y1 , y2 , . . ., y K .
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 25

S1 S2 S3 S4 ....... SK

t1 t2 t3 t4 tK

z1 z2 z3 z4 zK

o1 o2 o3 o4 oK

y1 y2 y3 y4 yK

n1 n2 n3 n4 nK

N1 N2 N3 N4 NK

FIGURE 2.6: DBN for a comprehensive generative model of speech from the phonological model to
distorted speech acoustics. Intermediate models include the target model, articulatory dynamic model, and
acoustic model. For clarity, only a one-tiered, rather than multitiered, phonological model is illustrated.
Explicitly added is the dependency of the parameters ((s )) of the articulatory dynamic model on the
phonological state

In Fig. 2.6, the dependency relationship for the new variables of distorted acoustic ob-
servation y(k) is specified, on the basis of observation equation 2.11 where o(k) is specified in
Eq. (5), by the following conditional PDF:

p(y(k) | o(k), n(k), h̄) = p  (y(k) − o(k) − h̄ − C log[I + exp(C−1 (n(k) − o(k) − h̄))]),
(2.12)
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

26 DYNAMIC SPEECH MODELS

where we assume that any inaccuracy in the parametric model of Eq. (2.11) can be repre-
sented by residual random noise [k]. This noise is assumed to be IID and zero-mean Gaus-
sian: N [(k); 0,  ]. This then specifies the conditional PDF of Eq. (2.12) to be Gaussian of
N [y(k); m,  ], where the mean vector m is the right-hand side of Eq. (2.11).
It is well known that the behavior of articulation and subsequent acoustics is subject
to modification under severe environmental distortions. This modification, sometimes called
“Lombard effect,” can take a number of possible forms, including articulatory target overshoot,
articulatory target shift, hyper-articulation or increased articulatory efforts by modifying the
temporal course of the articulatory dynamics. The Lombard effect has been very difficult to
represent in the conventional HMM framework since there is no articulatory representation or
any similar dynamic property therein. Given the generative model of speech described here that
explicitly contains articulatory variables, the Lombard effect can be naturally incorporated. Fig.
2.7 shows the DBN that incorporates Lombard effect in the comprehensive generative model of
speech. It is represented by the “feedback” dependency from the noise and h-distortion nodes to
the articulator nodes in the DBN. The nature of the feedback may be represented in the form of
“hyper-articulation,” where the “time constant” in the articulatory dynamic equation is reduced
to allow for more rapid attainment of the given articulatory target (which is sampled from the
target distribution). The feedback for Lombard effect may alternatively take the form of “target
overshoot,” where the articulatory dynamics exhibit oscillation around the articulatory target.
Finally, the feedback may take the form of “target elevation,” where the mean vector of the target
distribution is shifted further away from the target value of the preceding phonological state
compared with the situation when no Lombard effect occurs. Any of these three articulatory
behavior changes may result in enhanced discriminability among speech units under severe
environmental distortions, at the expense of greater articulatory efforts.

2.3.6 Piecewise Linearized Approximation for Articulatory-to-Acoustic Mapping

The nonlinearity h[z(k)] of Eq. (2.6) is a source of difficulty in developing efficient and effective
modeling learning algorithms. While the use of neural networks such as MLP or RBF as
described in the preceding subsection makes it possible to design such algorithms, a more
convenient strategy is to simplify the model by piecewise linearizing the nonlinear mapping
function h[z(k)] of Eq. (2.6). An extensive study based on nonlinear learning of the MLP-
based model can be found in [24]), where a series of approximations are required to complete
the algorithm development. Using piecewise linearization in the model would eliminate these
approximations.
After the model simplification, it is hoped that the piecewise linear methods will lead
to an adequate approximation to the nonlinear relationship between the hidden and observa-
tional spaces in the formulation of the dynamic speech model while gaining computational
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 27

S1 S2 S3 S4 ....... SK

t1 t2 t3 t4 tK

z1 z2 z3 z4 zK

o1 o2 o3 o4 oK

y1 y2 y3 y4 yK

n1 n2 n3 n4 nK

N1 N2 N3 N4 NK

FIGURE 2.7: DBN that incorporates the Lombard effect in the comprehensive generative model of
speech. The behavior of articulation is subject to modification (e.g., articulatory target overshoot or hyper-
articulation or increased articulatory efforts by shortening time constant) under severe environmental
distortions. This is represented by the “feedback” dependency from the noise nodes to the articulator
nodes in the DBN

effectiveness and efficiency in model learning. The most straightforward method is to use a
set of linear regression functions to replace the general nonlinear mapping in Eq. (2.6), while
keeping intact the target-directed, linear state dynamics of Eq. (2.4). That is, rather than using
one single set of linear-model parameters to characterize each phonological state, multiple sets
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

28 DYNAMIC SPEECH MODELS

of linear-model parameters are used. This gives rise to the mixture of linear dynamic model as
extensively studied in [84, 85].
This piecewise linearized dynamic speech model can be written succinctly in the following
state–space form (for a fixed phonological state s not shown for notational simplicity):

z(k + 1) = m z(k) + (I − m )tm + wm (k), (2.13)

o(k) = Ḣm ż(k) + vm (k), m = 1, 2, . . . , M, (2.14)

where Ḣm = [a | Hm ] is the expanded matrix by left appending vector a to matrix Hm , and
ż(k) = [1 | z(k) ] is the expanded vector in a similar manner. In the above equations, M is the
total number of mixture components in the model for each phonological state (e.g., phone). The
state noise and measurement noise, wm (k) and vm (k), are respectively modeled by uncorrelated,
IID, zero-mean, Gaussian processes with covariance matrices Qm and Rm . o represents the
sequence of acoustic vectors, o(1), o(2), . . . , o(k). . . , and the z represents the sequence of
hidden articulatory vectors, z(1), z(2), . . . , z(k), . . . .
The full set of model parameters for each phonological state (not indexed for clarity) are
 = (m , tm , Qm , Rm , Hm , for m = 1, 2, . . . , M).
It is important to impose the following mixture-path constraint on the above dynamic
system model: for each sequence of acoustic observations associated with a phonological state,
the sequence is forced to be produced from a fixed mixture component, m, in the model. This
means that the articulatory target for each phonological state is not permitted to switch from
one mixture component to another within the duration of the same segment. The constraint
is motivated by the physical nature of the dynamic speech model—the target that is correlated
with its phonetic identity is defined at the segment level, not at the frame level. Use of the
type of segment-level mixture is intended to represent the various sources of speech variability
including speakers’ vocal tract shape differences and speaking-habit differences, etc.
In Fig. 2.8 is shown the DBN representation for the piecewise linearized dynamic speech
model as a simplified generative model of speech where the nonlinear mapping from hidden
dynamic variables to acoustic observational variables is approximated by a piecewise linear rela-
tionship. The new, discrete random variable m is introduced to provide the “region” or mixture-
component index m to the piecewise linear mapping. Both the input and output variables that
are in a nonlinear relationship have now simultaneous dependency on m. The conditional PDFs
involving this new node are

p[o(k) | z(k), m] = N [o(k); Ḣm ż(k), Rm ], (2.15)

and

p[z(k + 1) | z(k), t(k), s, m] = N [z(k + 1); s ,m z(k) − (I − s ,m )t(k), Qm ], (2.16)

where k denotes the time frame and s denotes the phonological state.
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

A GENERAL MODELING AND COMPUTATIONAL FRAMEWORK 29

S1 S2 S3 S4 ....... SK

t1 t2 t3 t4 tK

z1 z2 z3 z4 zK

m1 m2 m3 m4 mK

o1 o2 o3 o4 oK

FIGURE 2.8: DBN representation for a mixture linear model as a simplified generative model of
speech where the nonlinear mapping from hidden dynamic variables to acoustic observational variables
is approximated by a piecewise linear relationship. The new, discrete random variable m is introduced to
provide “region” index to the piecewise linear mapping. Both the input and output variables that are in a
nonlinear relationship have now simultaneous dependency on m

2.4 SUMMARY
After providing general motivations and model design philosophy, technical detail of a multi-
stage statistical generative model of speech dynamics and its associated computational frame-
work based on DBN is presented in this chapter. We now summarize this model description.
Equations (2.4) and (2.6) form a special version of the switching state–space model appropriate
for describing the multilevel speech dynamics. The top-level dynamics occur at the discrete-state
phonology, represented by the state transitions of s with a relatively long time scale (roughly
about the duration of phones). The next level is the target (t) dynamics; it has the same time
scale and provides systematic randomness at the segmental level. At the level of articulatory
dynamics, the time scale is significantly shortened. This level represents the continuous-state
dynamics driven by the stochastic target process as the input. The state equation (2.4) explicitly
describes the dynamics in z, with index of s (which takes discrete values) implicitly representing
the phonological process of transitions among a set of discrete states, which we call “switching.”
At the lowest level is acoustic dynamics, where there is no phonological switching process.
Since the observation equation (2.6) is static, this simplified acoustic generation model assumes
that acoustic dynamics are a direct consequence of articulatory dynamics only. Improvement
of this model component that overcomes this simplification is unlikely until better modeling
P1: IML/FFX P2: IML
MOBK024-02 MOBK024-LiDeng.cls May 30, 2006 12:56

30 DYNAMIC SPEECH MODELS

techniques are developed for representing multiple time scales in the dynamic aspects of speech
acoustics.
Due to the generality of the DBN-based computational framework that we adopt, it
becomes convenient to extend the above generative model of speech dynamics one step further
from undistorted speech acoustics to distorted (or noisy) ones. We included this extension in
this chapter. Another extension that includes the changed articulatory behavior due to acoustic
distortion of speech is presented also within the same DBN-based computational framework.
Finally, we discussed piecewise linear approximation in the nonlinear articulatory-to-acoustic
mapping component of the overall model.
P1: IML/FFX P2: IML
MOBK024-03 MOBK024-LiDeng.cls May 16, 2006 14:4

31

CHAPTER 3

Modeling: From Acoustic Dynamics

to Hidden Dynamics

In Chapter 2, we described a rather general modeling scheme and the DBN-based computa-
tional framework for speech dynamics. Detailed implementation of the speech dynamic models
would vary depending on the trade-offs in modeling precision and mathematical/algorithm
tractability. In fact, various types of statistical models of speech beyond the HMM have already
been in the literature for sometime, although most of them have not been viewed from a unified
perspective as having varying degrees of approximation to the multistage speech chain. The
purpose of this chapter is to take this unified view in classifying and reviewing a wide variety of
current statistical speech models.

3.1 BACKGROUND AND INTRODUCTION

As we discussed earlier in this book, as a linguistic and physical abstraction, human speech pro-
duction can be functionally represented at four distinctive but correlated levels of dynamics. The
top level of the dynamics is symbolic or phonological. The multitiered linear sequence demon-
strates the discrete, time-varying nature of speech dynamics at the mental motor-planning
level of speech production. The next level of the dynamics is continuous-valued and asso-
ciated with the functional, “task” variables in speech production. At this level, the goal or
“task” of speech generation is defined, which may be either the acoustic goal such as vocal
tract resonances or formants, or the articulatory goal such as vocal-tract constrictions, or their
combination. It is at this level that each symbolic phonological unit is mapped to a unique set
of the phonetic parameters. These parameter is often called the correlate of the phonological
units. The third level of the dynamics occurs at the physiological articulators. Such articulatory
dynamics are a nonlinear transformation of the task dynamics. Finally, the last level of the
dynamics is the acoustic one, where speech “observations” are extracted from the speech signal.
They are often called acoustic observations or “feature” vectors in automatic speech recogni-
tion applications, and are called speech measurements in experimental phonetics and speech
science.
P1: IML/FFX P2: IML
MOBK024-03 MOBK024-LiDeng.cls May 16, 2006 14:4

32 DYNAMIC SPEECH MODELS

The review of several different types of computational dynamic models for speech in
this chapter will be organized in view of the above functional levels of speech dynamics. We
will classify the models into two main categories. In the first category are the models focusing
on the lowest, acoustic level of dynamics, which is also the most peripheral level for human
or computer speech perception. This class of models is often called the stochastic segment
model as is known through an earlier review paper [14]. The second category consists of what
is called the hidden dynamic model where the task dynamic and articulatory dynamic levels are
functionally grouped into a functional single-level dynamics. In contrast to the acoustic-dynamic
model, which represents coarticulation at the surface, observational level, the hidden dynamic
model explores a deeper, unobserved (hence “hidden”) level of the speech dynamic structure
that regulates coarticulation and phonetic reduction.

3.2 STATISTICAL MODELS FOR ACOUSTIC SPEECH DYNAMICS

Hidden Markov model (HMM) is the simplest type of the acoustic dynamic model in this
category. Stochastic segment models are a broad class of statistical models that generalize from
the HMM and that intend to overcome some shortcomings of the HMM such as the conditional
independent assumption and its consequences. As discussed earlier in this book, this assumption
is grossly unrealistic and restricts the ability of the HMM as an accurate generative model. The
generalization of the HMM by acoustic dynamic models is in the following sense: In an HMM,
one frame of speech acoustics is generated by visiting each HMM state, while a variable-length
sequence of speech frames is generated by visiting each “state” of a dynamic model. That is,
a state in the acoustic dynamic or stochastic segment model is associated with a “segment” of
acoustic speech vectors having a random sequence length.
Similar to an HMM, a stochastic segment model can be viewed as a generative process
for observation sequences. It is intended to model the acoustic feature trajectories and tem-
poral correlations that have been inadequately represented by an HMM. This is accomplished
by introducing new parameters that characterize the trajectories and the temporal correla-
tions.
From the perspective of the multilevel dynamics in the human speech process, the acoustic
dynamic model can be viewed as a highly simplified model—collapsing all three lower phonetic
levels of speech dynamics into one single level. As a result, the acoustic dynamic models have
difficulties in capturing the structure of speech coarticulation and reduction. To achieve high
performance in speech recognition, they tend to use many parallel (as opposed to hierarchi-
cal structured) parameters to model variability in acoustic dynamics, much like the strategies
adopted by the HMM.
A convenient way to understand a variety of acoustic dynamic models and their relation-
ships is to establish a hierarchy showing how the HMM is generalized by gradually relaxing the
P1: IML/FFX P2: IML
MOBK024-03 MOBK024-LiDeng.cls May 16, 2006 14:4

MODELING: FROM ACOUSTIC DYNAMICS TO HIDDEN DYNAMICS 33

modeling assumptions. Starting with a conventional HMM in this hierarchy, there are two main
classes of its extended or generalized models. Each of these classes further contains subclasses
of models. We describe this hierarchy below.

3.2.1 Nonstationary-State HMMs

This model class has also been called the trended HMM, constrained mean trajectory model,
segmental HMM, or stochastic trajectory model, etc., with minor variations according to
whether the parameters defining the trend functions or trajectories are random or not and
how their temporal properties are constrained. The trajectories for each state or segment are
sometimes normalized in time, especially when the linguistic unit associated with the state is
large (e.g., a word).
Given the HMM state s , the sample paths of most of these model types are explicitly
defined acoustic feature trajectories:

o(k) = gk (Λs ) + rs (k), (3.1)

where gk (Λs ) is the deterministic function of time frame k, parameterized by state-specific Λs ,

which can be either deterministic or random. And rs (k) is a state-specific stationary residual
signal.
The trend function gk (Λs ) in Eq. (3.1) varies with time (as indexed by k), and hence
describes acoustic dynamics. This is a special type of dynamics where no temporal recursion is
involved in characterizing the time-varying function. Throughout this book, we call this special
type of the dynamic function as a “trajectory,” or a kinematic function.
We now discuss further classification of the nonstationary-state or trended HMMs.

Polynomial Trended HMM

In this subset of the nonstationary-state HMMs, the trend function associated with each
HMM state is a polynomial function of time frames. Two common types of such models are as
follows:

• Observable polynomial trend functions: This is the simplest trended HMM where there
is no uncertainty in the polynomial coefficients Λs (e.g., [41, 55, 56, 86]).
• Random polynomial trend functions: The trend functions gk (Λs ) in Eq. (3.1) are stochas-
tic due to the uncertainty in polynomial coefficients Λs . Λs are random vectors in one
of the two ways: (1) Λs has a discrete distribution [87, 88] and (2) Λs has a continuous
distribution. In the latter case, the model is called the segmental HMM, where the
earlier versions have a polynomial order of zero [40, 89] and the later versions have an
order of one [90] or two [91].
P1: IML/FFX P2: IML
MOBK024-03 MOBK024-LiDeng.cls May 16, 2006 14:4

34 DYNAMIC SPEECH MODELS

Nonparametric Trended HMM
The trend function is determined by the training data after performing dynamic time warping
[92], rather than by any parametric form.

Observation-dependent Trended Function

In this rather recent nonstationary-state HMM, the trend function is designed in a special
way, where the parameters in Λs in the function gk (Λs ) of Eq. (3.1) are made dependent on
the observation vector o(k). The dependency is nonlinear, based on the posterior probability
computation [26].

3.2.2 Multiregion Recursive Models

Common to this model class is the recursive form in dynamic modeling of the region-dependent
time-varying acoustic feature vectors, where the “region” or state is often associated with a
phonetic unit. The most typical recursion is of the following linear form:

o(k) = Λs (1)o(k − 1) + · · · + Λs ( p)o(k − p) + rs (k), (3.2)

and the starting point of the recursion for each state s comes usually from the previous state’s
ending history.
The model expressed in Eq. (3.2) provides clear contrast to the trajectory or trended
models where the time-varying acoustic observation vectors are approximated as an explicit
temporal function of time. The sample paths of the model Eq. (3.2), on the other hand, are
piecewise, recursively defined stochastic time-varying functions. Further classification of this
model class is discussed below.

Autoregressive or Linear-predictive HMM

In this model, the time-varying function associated with each region (a Markov state) is defined
by linear prediction, or recursively defined autoregressive function. The work in [93] and that
in [94] developed this type of model having the state-dependent linear prediction performed on
the acoustic feature vectors (e.g., cepstra), with a first-order prediction and a second-order linear
prediction, respectively. The work in [95, 96] developed the model having the state-dependent
linear prediction performed on the speech waveforms. The latter model is also called the hidden
filter model in [95].

Dynamics Defined by Jointly Optimized Static and Delta Parameters

In this more recently introduced HMM version with recursively defined state-bound dy-
namics on acoustic feature vectors, the dynamics are in the form of joint static and delta
parameters [57, 97, 98]. The coefficients in the recursion are fixed for the delta parameters,
instead of being optimized as in the linear-predictive HMM. The optimized feature-vector
P1: IML/FFX P2: IML
MOBK024-03 MOBK024-LiDeng.cls May 16, 2006 14:4

MODELING: FROM ACOUSTIC DYNAMICS TO HIDDEN DYNAMICS 35

“trajectories” are obtained by joint use of static and delta model parameters. The results of the
constrained optimization provide an explicit relationship between the static and delta acoustic
features.

Nonlinear-predictive HMM
Several versions of nonlinear-predictive HMM have appeared in the literature, which generalize
the linear prediction in Eq. (3.2) to nonlinear prediction using neural networks (e.g., [99–101]).
In the model of [101], detailed statistical analysis was provided, proving that nonlinear prediction
with a short temporal order effectively produces a correlation structure over a significantly longer
temporal span.

Switching Linear Dynamic System Model

In this subclass of the multiregion recursive linear models, in addition to the use of the autoregres-
sive function that recursively defines the region-bound dynamics, which we call (continuous)-
state dynamics, a new noisy observation function is introduced. The actual effect of autore-
gression in this model is to smoothen the observed acoustic feature vectors. This model was
originally introduced in [102] for speech modeling.

3.3 STATISTICAL MODELS FOR HIDDEN SPEECH DYNAMICS

The various types of acoustic dynamic or stochastic segment models described in this chapter
generalize the HMM by generating a variable-length sequence of speech frames in each state,
overcoming the HMM’s assumption of local conditional independence. Yet the inconsistency
between the HMM assumptions and the properties of the realistic dynamic speech process goes
beyond this limitation. In acoustic dynamic models, the speech frames assigned to the same
segment/state have been modeled to be temporally correlated and the model parameters been
time-varying. However, the lengths of such segments are typically short. Longer-term correla-
tion across phonetic units, which provides dynamic structure responsible for coarticulation and
phonetic reduction, in a full utterance has not been captured.
This problem has been addressed by a class of more advanced dynamic speech models,
which we call hidden dynamic models. A hidden dynamic model exploits an intermediate level
of speech dynamics, functionally representing a combined system for speech motor control,
task dynamics, and articulatory dynamics. This intermediate level is said to be hidden since it is
not accessed directly from the speech acoustic data. The speech dynamics in this intermediate,
hidden level explicitly captures the long-contextual-span properties over the phonetic units
by imposing continuity constraints on the hidden dynamic variables internal to the acoustic
observation data. The constraint is motivated by physical properties of speech generation. The
constraint captures some key coarticulation and reduction properties in speech, and makes
the model parameterization more parsimonious than does the acoustic dynamic model where
P1: IML/FFX P2: IML
MOBK024-03 MOBK024-LiDeng.cls May 16, 2006 14:4

36 DYNAMIC SPEECH MODELS

modeling coarticulation requires a large number of free parameters. Since the underlying speech
structure represented by the hidden dynamic model links a sequence of segments via continuity
in the hidden dynamic variables, it can also be appropriately called the a super-segmental
model.
Differing from the acoustic dynamic models, the hidden dynamic models represent speech
structure by the hidden dynamic variables. Depending on the nature of these dynamic variables
in light of multilevel speech dynamics discussed earlier, the hidden dynamic models can be
broadly classified into

• articulatory dynamic model (e.g., [46, 54, 58, 59, 78, 79, 103, 104]);
• task-dynamic model (e.g., [105, 106]);
• vocal tract resonance (VTR) dynamic model (e.g., [24, 42, 48, 49, 84, 85, 107–112]);
• model with abstract dynamics (e.g., [42, 44, 107, 113]).

The VTR dynamics are a special type of task dynamics, with the acoustic goal or “task”
of speech production in the VTR domain. Key advantages of using VTRs as the “task” are their
direct correlation with the acoustic information, and the lower dimensionality in the VTR vector
compared with the counterpart hidden vectors either in the articulatory dynamic model or in
the task-dynamic model with articulatorily defined goal or “task” such as vocal tract constriction
properties.
As an alternative classification scheme, the hidden dynamic models can also be classified,
from the computational perspective, according to whether the hidden dynamics are represented
mathematically with temporal recursion or not. Like the acoustic dynamic models, the two
types of the hidden dynamic models in this classification scheme are reviewed here.

3.3.1 Multiregion Nonlinear Dynamic System Models

The hidden dynamic models in this first model class use the temporal recursion (k-recursion
via the predictive function gk in Eq. (3.3)) to define the hidden dynamics z(k). Each region, s ,
of such dynamics is characterized by the s -dependent parameter set Λs , with the “state noise”
denoted by ws (k). The memoryless nonlinear mapping function is exploited to link the hidden
dynamic vector z(k) to the observed acoustic feature vector o(k), with the “observation noise”
denoted by vs (k), and also parameterized by region-dependent parameters. The combined “state
equation” (3.3) and “observation equation” (3.4) form a general multiregion nonlinear dynamic
system model:

z(k + 1) = gk [z(k), Λs ] + ws (k), (3.3)

  
o(k ) = hk  [z(k ), Ωs  ] + vs  (k ). (3.4)
P1: IML/FFX P2: IML
MOBK024-03 MOBK024-LiDeng.cls May 16, 2006 14:4

MODELING: FROM ACOUSTIC DYNAMICS TO HIDDEN DYNAMICS 37


where subscripts k and k indicate that the functions g[·] and h[·] are time-varying and may
be asynchronous with each other. The subscripts s or s  denotes the dynamic region correlated
with phonetic categories.
Various simplified implementations of the above generic nonlinear system model have
appeared in the literature (e.g., [24, 33, 42, 45, 46, 59, 85, 108]). Most of these implementations
reduce the predictive function gk in the state equation (3.3) into a linear form, and use the
concept of phonetic targets as part of the parameters. This gives rise to linear target filtering
(by infinite impulse response or IIR filters) as a model for the hidden dynamics. Also, many of
these implementations use neural networks as the nonlinear mapping function hk [z(k), Ωs ] in
the observation equation (3.4).

3.3.2 Hidden Trajectory Models

The second type of the hidden dynamic models use trajectories (i.e., explicit functions of
time with no recursion) to represent the temporal evolution of the hidden dynamic variables
(e.g., VTR or articulatory vectors). This hidden trajectory model (HTM) differs conceptually
from the acoustic dynamic or trajectory model in that the articulatory-like constraints and
structure are captured in the HTM via the continuous-valued hidden variables that run across
the phonetic units. Importantly, the polynomial trajectories, which were shown to fit well
to the temporal properties of cepstral features [55, 56], are not appropriate for the hidden
dynamics that require realistic physical constraints of segment-bound monotonicity and target-
directedness. One parametric form of the hidden trajectory constructed to satisfy both these
constraints is the critically damped exponential function of time [33, 114]. Another parametric
form of the hidden trajectory, which also satisfies these constraints but with more flexibility to
handle asynchrony between segment boundaries for the hidden trajectories and for the acoustic
features, has been developed more recently [109,112,115,116] based on finite impulse response
(FIR) filtering of VTR target sequences. In Chapter 5, we provide a systematic account of this
model, synthesizing and expanding the earlier descriptions of this work in [109, 115, 116].

3.4 SUMMARY
This chapter serves as a bridge between the general modeling and computational framework
for speech dynamics (Chapter 2) and Chapters 4 and 5 on detailed descriptions of two specific
implementation strategies and algorithms for hidden dynamic models. The theme of this chapter
is to move from the relatively simplistic view of dynamic speech modeling confined within the
acoustic stage to the more realistic view of multistage speech dynamics with an intermediate
hidden dynamic layer between the phonological states and the acoustic dynamics. The latter,
with appropriate constraints in the form of the dynamic function, permits a representation
of the underlying speech structure responsible for coarticulation and speaking-effort-related
P1: IML/FFX P2: IML
MOBK024-03 MOBK024-LiDeng.cls May 16, 2006 14:4

38 DYNAMIC SPEECH MODELS

reduction. This type of structured modeling is difficult to accomplish by acoustic dynamic
models with no hidden dynamic layer, unless highly elaborate model parameterization is carried
out. In Chapter 5, we will show an example where a hidden trajectory model can be simplified
to an equivalent of an acoustic trajectory model whose trajectory parameters become long-
span context-dependent via a structured means and delicate parameterization derived from the
construction of the hidden trajectories.
Guided by this theme, in this chapter we classify and review a rather rich body of literature
on a wide variety of statistical models of speech, starting with the traditional HMM [4] as
the most primitive model. Two major classes of the models, acoustic dynamic models and
hidden dynamic models, respectively, are each further classified into subclasses based on how the
dynamic functions are constructed. When explicit temporal functions are constructed without
recursion, then we have classes of “trajectory” models. The trajectory models and recursively
defined dynamic models can achieve a similar level of modeling accuracy but they demand very
different algorithm development for model parameter learning and for speech decoding. Each
of these two classes (acoustic vs. hidden dynamic) and two types (trajectory vs. recursive) of the
models simplifies, in different ways, the DBN structure as the general computational framework
for the full multistage speech chain (Chapter 2).
In the remaining two chapters, we select two types of hidden dynamic models of speech for
their detailed exposition, one with and another without recursion in defining the hidden dynamic
variables. The exposition will include the implementation strategies (discretization of the hidden
dynamic variables or otherwise) and the related algorithms for model parameter learning and
model scoring/decoding. The implementation strategy with discretization of recursively defined
hidden speech dynamics will be covered in Chapter 4, and the strategy using hidden trajectories
(i.e., explicit temporal functions) with no discretization will be discussed in Chapter 5.
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

39

CHAPTER 4

Models with Discrete-Valued Hidden

Speech Dynamics

In this chapter, we focus on a special type of hidden dynamic models where the hidden dynamics
are recursively defined and where these hidden dynamic values are discretized. The discretization
or quantization of the hidden dynamics causes an approximation to the original continuous-
valued dynamics as described in the earlier chapters but it enables an implementation strategy
that can take direct advantage of the forward–backward algorithm and dynamic programming
in model parameter learning and decoding. Without discretization, the parameter learning and
decoding problems would be typically intractable (i.e., the computation cost would increase
exponentially with time). Under different kinds of model implmentation schemes, other types
of approximation will be needed and one type of the approximation in this case will be detailed
in Chapter 5.
This chapter is based on the materials published in [110, 117], with reorganization,
rewriting, and expansion of these materials so that they naturally fit as an integral part of this
book.

4.1 BASIC MODEL WITH DISCRETIZED HIDDEN DYNAMICS

In the basic model presented in this section, we assume discrete-time, first-order hidden dy-
namics in the state equation and linearized mapping from the hidden dynamic variables to the
acoustic observation variables in the observation equation. Before discretizing hidden dynam-
ics, the first-order dynamics in a scalar form have the following form (which was discussed in
Chapter 2 with a vector form):

xt = r s xt−1 + (1 − r s )Ts + wt (s ), (4.1)

where state noise wt ∼ N(wk ; 0, Bs ) is assumed to be IID, zero-mean Gaussian with phonolog-
ical state (s )-dependent precision (inverse of variance) Bs . The linearized observation equation
is

o t = Hs xt + h s + vt , (4.2)
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

40 DYNAMIC SPEECH MODELS

where observation noise vk ∼ N(vk ; 0, Ds ) is assumed to be IID, zero-mean Gaussian with
precision Ds .
We now perform discretization or quantization on hidden dynamic variable xt . For sim-
plicity in illustration, we use scalar hidden dynamics most of the times in this chapter (except
Section 4.2.3) where scalar quantization is carried out, and let C denote the total number of
discretization/quantization levels. (For the more realistic, multidimensional hidden dynamic
case, C would be the total number of cells in the vector-quantized space.) In the following
derivation of the EM algorithm for parameter learning, we will use variable xt [i] or it to denote
the event that at time frame t the state variable (or vector) xt takes the mid-point (or centroid)
value associated with the ith discretization level in the quantized space.
We now describe this basic model with discretized hidden dynamics in an explicit prob-
abilistic form and then derive and present a maximum-likelihood (ML) parameter estimation
technique based on the Expectation-Maximization (EM) algorithm. The background infor-
mation on ML and EM can be found in of [9], [Part I, Ch. 5, Sec. 5.6].

4.1.1 Probabilistic Formulation of the Basic Model

Before discretization, the basic model that consists of Eqs. (4.1) and (4.2) can be equivalently
written in the following explicit probabilistic form:

p(xt | xt−1 , s t = s ) = N(xt ; r s xt−1 + (1 − r s )Ts , Bs ), (4.3)

p(o t | xt , s t = s ) = N(o t ; Hs xt + h s , Ds ). (4.4)

And we also have the transition probability for the phonological states:

p(s t = s | s t−1 = s  ) = πs  s .

Then the joint probability can be written as


N
p(s 1N , x1N , o 1N ) = πs t−1 s t p(xt | xt−1 , s t ) p(o t | xt , s t = s ),
t=1

where N is the total number of observation data points in the training set.
After discretization of hidden dynamic variables, Eqs. (4.3) and (4.4) are approximated
as

p(xt [i] | xt−1 [ j ], s t = s ) ≈ N(xt [i]; r s xt−1 [ j ] + (1 − r s )Ts , Bs ), (4.5)

and

p(o t | xt [i], s t = s ) ≈ N(o t ; Hs xt [i] + h s , Ds ). (4.6)

P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 41

4.1.2 Parameter Estimation for the Basic Model: Overview
For carrying out the EM algorithm for parameter estimation of the above discretized model,
we first establish the auxiliary function, Q. Then we simplify the Q function into a form that
can be optimized in a closed form.
According to the EM theory, the auxiliary objective function Q is the conditional expecta-
tion of logarithm of the joint likelihood of all hidden and observable variables. The conditioning
events are all observation sequences in the training data:

o 1N = o 1 , o 2 , . . . , o t , . . . , o N ,
And the expectation is taken over the posterior probability for all hidden variable sequences:

x1N = x1 , x2 , . . . , xt , . . . , x N ,
and

s 1N = s 1 , s 2 , . . . , s t , . . . , s N ,
This gives (before discretization of the hidden dynamic variables):
    
Q= ··· ··· ··· ··· p(s 1N , x1N | o 1N ) log p(s 1N , x1N , o 1N )d x1 · · · d xt · · · d x N ,
s1 st sN x1 xt xN
(4.7)
where the summation for each phonological state s is from 1 to S (the total number of distinct
phonological units).
After discretizing xt into xt [i], the objective function of Eq. (4.7) is approximated by
    
Q≈ ··· ··· ··· ··· p(s 1N , i1N | o 1N ) log p(s 1N , i1N , o 1N ), (4.8)
s1 st sN i1 it iN

where the summation for each discretization index i is from 1 to C.

We now describe details of the E-step and m-Step in the EM algorithm.

4.1.3 EM Algorithm: The E-Step

The following outlines the simplification steps for the objective function of Eq. (4.8). Let us de-
      
note the sequence summation s 1 · · · s t · · · s N by s 1N , and summation i1 · · · it · · · i N

by i1N . Then we rewrite Q in Eq. (4.8) as

Q(r s , Ts , Bs , Hs , h s , Ds ) ≈ p(s 1N , i1N | o 1N ) log p(s 1N , i1N , o 1N ) (4.9)
s 1N i1N
 
= p(s 1N , i1N |o 1N ) log p(o 1N | s 1N , i1N ) + p(s 1N , i1N | o 1N ) log p(s 1N , i1N ),
sN i1N sN i1N
1  1 
Q o (Hs ,h s ,Ds ) Q x (r s ,Ts ,Bs )
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

42 DYNAMIC SPEECH MODELS

where

p(s 1N , i1N ) = πs t−1 s t N(xt [i]; r s t xt−1 [ j ] + (1 − r s t )Ts t , Bs t ),
t

and

p(o 1N | s 1N , i1N ) = N(o t ; Hs t xt [i] + h s t , Ds t ).
t

In these equations, discretization indices i and j denote the hidden dynamic values taken at
time frames t and t − 1, respectively. That is, s t = i, s t−1 = j .
We first compute Q o (omitting constant −0.5d log(2π) that is irrelevant to optimization):
 
N
2
Q o = 0.5 p(s 1N , i1N | o 1N ) log |Ds t | − Ds t o t − Hs t xt [i] − h s t
s 1N i1N t=1

 C 
S   
N 
2
= 0.5 p(s 1N , i1N | o 1N ) log |Ds t | − Ds t o t − Hs t xt [i] − h s t δs t s δit i
s =1 i=1 s 1N i1N t=1


S  N  
C  
= 0.5 p(s 1N , i1N | o 1N )δs t s δit i log |Ds | − Ds (o t − Hs xt [i] − h s )2 .
s =1 i=1 t=1 s 1N i1N

Noting that

p(s 1N , i1N | o 1N )δs t s δit i = p(s t = s , it = i | o 1N ) = γt (s , i),
s 1N i1N

we obtain the simplified form of


S 
N 
C
Q o (Hs , h s , Ds ) = 0.5 γt (s , i) log |Ds | − Ds (o t − Hs xt [i] − h s )2 . (4.10)
s =1 t=1 i=1

Similarly, after omitting optimization-independent constants, we have

 
N
2
Q x = 0.5 p(s 1N , i1N | o 1N ) log |Bs t | − Bs t xt [i] − r s t xt−1 [ j ] − (1 − r s t )Ts t
s 1N i1N t=1


S  C 
C  
= 0.5 p(s 1N , i1N | o 1N )
s =1 i=1 j =1 s 1N i1N

N 
2
× log |Bs t | − Bs t xt [i] − r s t xt−1 [ j ] − (1 − r s t )Ts t δs t s δit i δit−1 j
t=1
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 43


S 
C  N  
C  
= 0.5 p(s 1N , i1N | o 1N )δs t s δit i δit−1 j
s =1 i=1 j =1 t=1 s 1N i1N

× log |Bs | − Bs (xt [i] − r s xt−1 [ j ] − (1 − r s )Ts )2 .

Now noting

p(s 1N , i1N | o 1N )δs t s δit i δit−1 j = p(s t = s , it = i, it−1 = j | o 1N ) = ξt (s , i, j ),
s 1N i1N

we obtain the simplified form of


S 
N 
C 
C

Q x (r s , Ts , Bs ) = 0.5 ξt (s , i, j ) log |Bs |
s =1 t=1 i=1 j =1

−Bs (xt [i] − r s xt−1 [ j ] − (1 − r s )Ts )2 , (4.11)

Note that large computational saving can be achieved by limiting the summations in
Eq. (4.11) for i, j based on the relative smoothness of hidden dynamics. That is, the range of
i, j can be limited such that |xt [i] − xt−1 [ j ]| < Th, where Th is empirically set threshold value
that controls the computation cost and accuracy.
In Eqs. (4.11) and (4.10), we used ξt (s , i, j ) and γt (s , i) to denote the single-frame
posteriors of

ξt (s , i, j ) ≡ p(s t = s , xt [i], xt−1 [ j ] | o 1N ),

and

γt (s , i) ≡ p(s t = s , xt [i] | o 1N ).

These can be computed efficiently using the generalized forward–backward algorithm (part of
the E-step), which we describe below.

4.1.4 A Generalized Forward–Backward Algorithm

The only quantities that need to be determined in simplified auxiliary function Q = Q o + Q x
as in Eqs. (4.9)–(4.11) are the two frame-level posteriors ξt (s , i, j ) and γt (s , i), which we
compute now in order to complete the E-step in the EM algorithm.

Generalized α(s t , it ) Forward Recursion

The generalized forward recursion discussed here uses a new definition of the variable

αt (s , i) ≡ p(o 1t , s t = s , it = i).
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

44 DYNAMIC SPEECH MODELS

The generalization of the standard forward–backward algorithm for HMM in any standard
textbook on speech recognition is by including additional discrete hidden variables related to
hidden dynamics.
For notational convenience, we use α(s t , it ) to denote αt (s , i) below. The forward recursive
formula is

S 
C
α(s t+1 , it+1 ) = α(s t , it ) p(s t+1 , it+1 |s t , it ) p(o t+1 |s t+1 , it+1 ). (4.12)
s t =1 it =1

Proof of Eq. (4.12):

α(s t+1 , it+1 ) ≡ p(o 1t+1 , s t+1 , it+1 )


= p(o 1t , o t+1 , s t+1 , it+1 , s t , it )
st it

= p(o t+1 , s t+1 , it+1 | o 1t , s t , it ) p(o 1t , s t , it )
st it

= p(o t+1 , s t+1 , it+1 | s t , it )α(s t , it )
st it

= p(o t+1 | s t+1 , it+1 , s t , it ) p(s t+1 , it+1 | s t , it )α(s t , it )
st it

= p(o t+1 | s t+1 , it+1 ) p(s t+1 , it+1 | s t , it )α(s t , it ). (4.13)
st it

In Eq. (4.12), p(o t+1 | s t+1 , it+1 ) is determined by the observation equation:

p(o t+1 | s t+1 = s , it+1 = i) = N(o t+1 ; Hs xt+1 [i] + h s , Ds ),

and p(s t+1 , it+1 | s t , it ) is determined by the state equation (with order one) and the switching
Markov chain’s transition probabilities:

p(s t+1 = s , it+1 = i | s t = s  , it = i  ) ≈ p(s t+1 = s | s t = s  ) p(it+1 = i | it = i  )

= πs t−1 s t p(it+1 = i | it = i  ). (4.14)

Generalized γ(s t , it ) Backward Recursion

Rather than performing backward β recursion and then combining the α and β to obtain
the single-frame posterior as for the conventional HMM, a more memory-efficient technique
can be used for backward recursion, which directly computes the single-frame posterior. For
notational convenience, we use γ (s t , it ) to denote γt (s , i) below.
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 45

The development of the generalized γ (s t , it ) backward recursion for the first-order state
equation proceeds as follows:

γ (s t , it ) ≡ p(s t , it | o 1N )

= p(s t , it , s t+1 , it+1 | o 1N )
s t+1 it+1

= p(s t , it , s t+1 , it+1 | o 1N ) p(s t+1 , it+1 | o 1N )
s t+1 it+1

= p(s t , it , s t+1 , it+1 | o 1t )γ (s t+1 , it+1 )
s t+1 it+1

  p(s t , it , s t+1 , it+1 , o t )

= 1
γ (s t+1 , it+1 ) (Bayes rule)
s t+1 it+1 p(s t+1 , i t+1 , o t
1 )
 p(s t , it , s t+1 , it+1 , o 1t )
=   t γ (s t+1 , i t+1 )
s t+1 it+1 st it p(s t , i t , s t+1 , i t+1 , o 1 )

 p(s t , it , o 1t ) p(s t+1 , it+1 | s t , it , o 1t )

=   t γ (s t+1 , i t+1 )
it p(s t , i t , o 1 ) p(s t+1 , i t+1 | s t , i t , o 1 )
t
s t+1 it+1 st

 α(s t , it ) p(s t+1 , it+1 | s t , it )

=   γ (s t+1 , it+1 ), (4.15)
s t+1 it+1 st it α(s t , i t ) p(s t+1 , i t+1 | s t , i t )

where the last step uses conditional independence, and where α(s t , it ) and p(s t+1 , it+1 | s t , it )
on the right-hand side of Eq. (4.15) have been computed already in the forward recursion.
Initialization for the above γ recursion is γ (s N , i N ) = α(s N , i N ), which will be equal to 1 for
the left-to-right model of phonetic strings.
Given this result, ξt (s , i, j ) can be computed directly using α(s t , it ) and γ (s t , it ). Both of
them are already computed from the forward–backward recursions described above.
Alternatively, we can compute β generalized recursion (not discussed here) and then
combine αs and βs to obtain γt (s , i) and ξt (s , i, j ).

4.1.5 EM Algorithm: The M-Step

Given the results of the E-step described above where the frame-level posteriors are computed
efficiently by the generalized forward–backward algorithm, we now derive the reestimation
formulas, as the M-step in the EM algorithm, by optimizing the simplified auxiliary function
Q = Q o + Q x as in Eqs. (4.9), (4.10) and (4.11).
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

46 DYNAMIC SPEECH MODELS

Reestimation for the Hidden-to-Observation Mapping Parameters Hs and h s
Taking partial derivatives of Q o in Eq. (4.10) with respect to Hs and h s , respectively, and setting
them to zero, we obtain:

∂ Q o (Hs , h s , Ds ) N  C
= −Ds γt (s , i){o t − Hs xt [i] − h s } = 0, (4.16)
∂h s t=1 i=1

and
∂ Q o (Hs , h s , Ds ) N  C
= −Ds γt (s , i){o t − Hs xt [i] − h s }xt [i] = 0. (4.17)
∂ Hs t=1 i=1

These can be rewritten as the standard linear system of equations:

U Ĥs + V1 ĥ s = C1 , (4.18)
V2 Ĥs + U ĥ s = C2 , (4.19)

where

N 
C
U= γt (s , i)xt [i], (4.20)
t=1 i=1
V1 = N, (4.21)
N 
C
C1 = γt (s , i)o t , (4.22)
t=1 i=1
N  C
V2 = γt (s , i)xt2 [i], (4.23)
t=1 i=1
N  C
C2 = γt (s , i)o t xt [i]. (4.24)
t=1 i=1

The solution is
   −1  
Ĥs U V1 C1
= . (4.25)
ĥ s V2 U C2

Reestimation for the Hidden Dynamic Shaping Parameter rs

Taking partial derivative of Q x in Eq. (4.11) with respect to r s and setting it to zero, we obtain
the reestimation formula of
∂ Q x (r s , Ts , Bs ) N  C  C
= −Bs ξt (s , i, j ) (4.26)
∂r s t=1 i=1 j =1

× xt [i] − r s xt−1 [ j ] − (1 − r s )Ts xt−1 [ j ] − Ts = 0.

P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 47

Solving for r s , we have
 −1
N  C 
C
r̂ s = ξt (s , i, j )(Ts − xt−1 [ j ])2

t=1 i=1 j =1
 

N 
C 
C
× ξt (s , i, j )(Ts − xt−1 [ j ])(Ts − xt [i]) , (4.27)
t=1 i=1 j =1

where we assume that all other model parameters are fixed.

It is interesting to note from the above that when xt is monotonically moving (on average)
toward the target Ts (i.e., no target overshooting), the reestimate of r s is guaranteed to be positive,
as it should be.

Reestimation for the Hidden Dynamic Target Parameter Ts

Similarly, taking partial derivative of Q x in Eq. (4.11) with respect to Ts and setting it to zero,
we obtain the reestimation formula of
∂ Q x (r s , Ts , Bs ) N  C  C
= −Bs ξt (s , i, j ) xt [i] − r s xt−1 [ j ] − (1 − r s )Ts (1 − r s ) = 0.
∂ Ts t=1 i=1 j =1
(4.28)
Solving for Ts , we obtain

1  N  C  C
T̂s = ξt (s , i, j ) xt [i] − r s xt−1 [ j ] . (4.29)
1 − r s t=1 i=1 j =1

Intuitions behind the target estimate above are particularly obvious.

Reestimation for the Noise Precisions Bs and Ds

Setting

∂ Q x (r s , Ts , Bs ) N  C  C
= 0.5 ξt (s , i, j ) Bs−1 − (xt [i] − r s xt−1 [ j ] − (1 − r s )Ts )2 = 0,
∂ Bs t=1 i=1 j =1

we obtain the state noise variance reestimate of

 N C C 2
−1 t=1 i=1 j =1 ξt (s , i, j ) [xt [i] − r̂ s xt−1 [ j ] − (1 − r̂ s )Ts ]
B̂s =  N C C . (4.30)
t=1 i=1 j =1 ξ t (s , i, j )

Similarly, setting

∂ Q o (Hs , h s , Ds ) N  C
= 0.5 γt (s , i) Ds−1 − (o t − Hs xt [i] − h s )2 = 0,
∂ Ds t=1 i=1
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

48 DYNAMIC SPEECH MODELS

we obtain the observation noise variance reestimate of
 N C
γt (s , i) [o t − Hs xt [i] − h s ]2
D̂s = t=1 i=1 N C . (4.31)
t=1 i=1 γt (s , i)

4.1.6 Decoding of Discrete States by Dynamic Programming

After the parameters of the basic model are estimated using the EM algorithm described above,
estimation of discrete phonological states and of the quantized hidden dynamic variables can be
carried out jointly. We call this process “decoding.” Estimation of the phonological states is the
problem of speech recognition, and estimation of the hidden dynamic variables is the problem
of tracking hidden dynamics. For large vocabulary speech recognition, aggressive pruning and
careful design of data structures will be required (which is not described in this book).
Before describing the decoding algorithm, which is aimed at finding the best single joint
state and quantized hidden dynamic variable sequences (s 1N , i1N ) for a given observation sequence
o 1N , let us define the quantity

δt (s , i) = max P (o 1t , s 1t−1 , i1t−1 , s t = s , xt [i])

s 1 ,s 2 ,...,s t−1 ,i1 ,i2 ,...,it−1

= max P (o 1t , s 1t−1 , i1t−1 , s t = s , it = i). (4.32)

s 1t−1 ,i1t−1

Note that each δt (s , i) defined here is associated with a node in a three-dimensional trellis
diagram. Each increment in time corresponds to reaching a new stage in dynamic programming
(DP). At the final stage t = N, we have the objective function of δ N (s , i) that is accomplished
via all the previous stages of computation for t ≤ N − 1. On the basis of the DP optimality
principle, the optimal (joint) partial likelihood at the processing stage of t + 1 can be computed
using the following DP recursion:

δt+1 (s , i) = max
 
δt (s  , i  ) p(s t+1 = s , it+1 = i | s t = s  , it = i  ) p(o t+1 | s t+1 = s , it+1 = i)
s ,i
≈ max
 
δt (s  , i  ) p(s t+1 = s | s t = s  ) p(it+1 = i | it = i  ) p(o t+1 | s t+1 = s , it+1 = i)
s ,i
= max
 
δt (s  , i  )πs  s N(xt+1 [i]; r s  xt [ j ] + (1 − r s  )Ts  , Bs  )
s ,i , j
×N(o t+1 ; Hs xt+1 [i] + h s , Ds ), (4.33)

for all states s and for all quantization indices i. Each pair of (s , i) at this processing stage is
a hypothesized “precursor” node in the global optimal path. All such nodes except one will be
eventually eliminated after the backtracking operation. The essence of DP used here is that we
only need to compute the quantities of δt+1 (s , i) as individual nodes in the trellis, removing
the need to keep track of a very large number of partial paths from the initial stage to the
current (t + 1)th stage, which would be required for the exhaustive search. The optimality is
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 49

guaranteed, due to the DP optimality principle, with the computation only linearly, rather than
geometrically, increasing with the length N of the observation data sequence o 1N .

4.2 EXTENSION OF THE BASIC MODEL

The preceding section presented details of the basic hidden dynamic model where the discretized
state equation takes the simplest first-order recursive form and where the observation equation
also takes the simplest linear form responsible for mapping from hidden dynamic variables to
acoustic observation variables. We now present an extension of this discretized basic model.
First, we will extend the state equation of the basic model from first-order dynamics to second-
order dynamics so as to improve the modeling accuracy. Second, we will extend the observation
equation of the basic model from the linear form to a nonlinear form of the mapping function
from the discretized hidden dynamic variables to (nondiscretized or continuous-valued) acoustic
observation variables.

4.2.1 Extension from First-Order to Second-Order Dynamics

In this first step of extension of the basic model, we change from the first-order state equation
(Eq. (4.1)):

xt = r s xt−1 + (1 − r s )Ts + wt (s ),

to the new second-order state equation

xt = 2r s xt−1 − r s2 xt−2 + (1 − r s )2 Ts + wt (s ). (4.34)

Here, like the first-order state equation, state noise wk ∼ N(wk ; 0, Bs ) is assumed to be IID
zero-mean Gaussian with state (s )-dependent precision Bs . And again, Ts is the target parameter
that serves as the “attractor” drawing the time-varying hidden dynamic variable toward it within
each phonological unit denoted by s .
It is easy to verify that this second-order state equation, as for the first-order one, has the
desirable properties of target directedness and monotonicity. However, the trajectory implied
by the second-order recursion is more realistic than that by the earlier first-order one. The
new trajectory has critically damped trajectory shaping, while the first-order trajectory has
exponential shaping. Detailed behaviors of the respective trajectories are controlled by the
parameter r s in both the cases. For analysis of such behaviors, see [33, 54].
The explicit probabilistic form of the state equation (4.34) is

p(xt | xt−1 , xt−2 , s t = s ) = N(xt ; 2r s xt−1 − r s2 xt−2 + (1 − r s )2 Ts , Bs ). (4.35)

Note the conditioning event is both xt−1 and xt−2 , instead of just xt−1 as in the first-order case.
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

50 DYNAMIC SPEECH MODELS

After discretization of the hidden dyanmic variables xt , xt−1 , and xt−2 , Eq. (4.35) turns
into an approximate form:
p(xt [i] | xt−1 [ j ], xt−2 [k], s t = s ) ≈ N(xt [i]; 2r s xt−1 [ j ] − r s2 xt−2 [k] + (1 − r s )2 Ts , Bs ).
(4.36)

4.2.2 Extension from Linear to Nonlinear Mapping

The second step of extension of the basic model involves changing from the linear form of the
observation equation
o t = Hs xt + h s + vt ,
to the new nonlinear form
o t = F(xt ) + h s + vt (s ), (4.37)
where the output of nonlinear predictive or mapping function F(xt ) is the acoustic measurement
that can be computed directly from the speech waveform. The expression h s + vt (s ) is the
prediction residual, where h s is the state-dependent mean and the observation noise vk (s ) ∼
N(vk ; 0, Ds ) is an IID, zero-mean Gaussian with precision Ds . The phonological unit or state
s in h s may be further subdivided into several left-to-right subunit states. In this case, we can
treat all the state labels s as the subphone states but tie the subphone states in the state equation
so that the sets of Ts , r s , Bs are the same for a given phonological unit. This will simplify
the exposition of the model in this section without having to distinguish the state from the
“substate” and we will use the same label s to denote both. The nonlinear function F(xt ) may
be made phonological-unit-dependent to increase the model discriminability (as in [24]). But
for simplicity, we assume that in this chapter it is independent of phonological units.
Again, we rewrite Eq. (4.37) in an explicit probabilistic form of

p(o t | xt , s t = s ) = N(o t ; F(xt ) + h s , Ds ). (4.38)

After discretizing the hidden dyanmic variable xt , the observation equation (4.38) is approxi-
mated by
p(o t | xt [i], s t = s ) ≈ N(o t ; F(xt [i]) + h s , Ds ). (4.39)
Combining this with Eq. (4.35), we have the joint probability model:

N
p(s 1N , x1N , o 1N ) = πs t−1 s t p(xt | xt−1 , xt−2 , s t ) p(o t | xt , s t = s )
t=1

N
≈ πs t−1 s t N(x[it ]; 2r s x[it−1 ] − r s2 x[it−2 ] + (1 − r s )2 Ts , Bs )
t=1
×N(o t ; F(x[it ]) + h s , Ds ), (4.40)
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 51

where it , it−1 , and it−2 denote the discretization indices of the hidden dynamic variables at time
frames t, t − 1, and t − 2, respectively.

4.2.3 An Analytical Form of the Nonlinear Mapping Function

The choice of the functional form of F(xt ) in Eq. (4.38) is critical for the success of the model
in applications. In Chapter 2, we discussed the use of neural network functions (MLP and RBF,
etc.) as well as the use of piecewise linear functions to represent or approximate the generally
nonlinear function responsible for mapping from the hidden dynamic variables to acoustic
observation variables. These techniques, while useful as shown in [24, 84, 85, 108, 118], either
require a large number of parameters to train, or necessitate crude approximation as needed for
carrying out parameter estimation algorithm development.
In this section, we will present a specific form of the nonlinear function of F(x) that
contains no free parameters and that after discretizing the input argument x invokes no further
approximation in developing and implementing the EM-based parameter estimation algorithm.
The key to developing this highly desirable form of the nonlinear function is to endow the hidden
dynamic variables with their physical meaning. In this case, we let the hidden dynamic variables
be vocal tract resonances (VTRs, and sometimes called formants) including both resonance
frequencies and bandwidths. Then, under reasonable assumptions, we can derive an explicit
nonlinear functional relationship between the hidden dynamic variables (in the form of VTRs)
and the acoustic observation variables in the form of linear cepstra [5]. We now describe this
approach in detail.

Definition of Hidden Dynamic Variables and Related Notations

Let us define the hidden dynamic variables for each frame of speech as the 2K -dimensional
vector of VTRs. It consists of a set of P resonant frequencies f and corresponding bandwidths
b, which we denote as
 
f
x= ,
b

where
⎛ ⎞ ⎛ ⎞
f1 b1
⎜ ⎟ ⎜ ⎟
⎜ f2 ⎟ ⎜ b2 ⎟
f=⎜
⎜ .. ⎟
⎟ and b=⎜ ⎟
⎜ .. ⎟ .
⎝ . ⎠ ⎝ . ⎠
fP bP
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

52 DYNAMIC SPEECH MODELS

We desire to establish a memoryless mapping relationship between the VTR vector x and
an acoustic measurement vector o:

o ≈ F(x).

Depending on the type of the acoustic measurements as the output in the mapping function,
closed-form computation for F(x) may be impossible, or its in-line computation may be too
expensive. To overcome these difficulties, we may quantize each dimension of x over a range
of frequencies or bandwidths, and then compute C(x) for every quantized vector value of x.
This will be made especially effective when a closed form of the nonlinear function can be
established. We will next show that when the output of the nonlinear function becomes linear
cepstra, a closed form can be easily derived.

Derivation of a Closed-form Nonlinear Function from VTR to Cepstra

Consider an all-pole model of speech, with each of its poles represented as a frequency–
bandwidth pair ( f p , b p ). Then the corresponding complex root is given by [119]
bp fp bp fp
−π + j 2π −π − j 2π
zp = e f samp f samp
, and z∗p = e f samp f samp
, (4.41)

where f samp is the sampling frequency. The transfer function with P poles and a gain of G is

P
1
H(z) = G . (4.42)
p=1
(1 − z p z−1 )(1 − z∗p z−1 )

Taking logarithm on both sides of Eq. (4.42), we obtain


P 
P
log H(z) = log G − log(1 − z p z−1 ) − log(1 − z∗p z−1 ). (4.43)
p=1 p=1

Now using the well-known infinite series expansion formula


∞
vn
log(1 − v) = − , |v| ≤ 1,
n=1
n

and with v = z p z−1 , we obtain

 

P 
∞ zn z−n 
P 
∞ z∗n z−n 
∞ 
P zn + z∗n
z−n .
p p p p
log H(z) = log G + + = log G +
p=1 n=1
n p=1 n=1
n n=1 p=1
n
(4.44)
Comparing Eq. (4.44) with the definition of the one-sided z-transform,

∞ 
∞
C(z) = c n z−n = c 0 + c n z−n ,
n=0 n=1
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 53

we immediately see that the inverse z-transform of log H(z) in Eq. (4.44), which by definition
is the linear cepstrum, is


P zn + z∗n
p p
cn = , n > 0, (4.45)
p=1
n

and c 0 = log G.
Using Eq. (4.41) to expand and simplify Eq. (4.45), we obtain the final form of the
nonlinear function (for n > 0):

1 P bp fp bp fp
cn = e −πn fs + j 2πn fs + e −πn fs − j 2π n fs
n p=1

1 P bp fp fp
= e −π n fs e j 2π n fs + e − j 2π n fs
n p=1
        
1 P b
−π n fps fp fp fp fp
= e cos 2πn + j sin 2πn + cos 2π n − j sin 2πn
n p=1 fs fs fs fs
 
2 P b
−π n fps fp
= e cos 2πn . (4.46)
n p=1 fs

Here, c n constitutes each of the elements in the vector-valued output of the nonlinear
function F(x).

Illustrations of the Nonlinear Function

Equation (4.46) gives the decomposition property of the linear cepstrum—it is a sum of the
contributions from separate resonances without interacting with each other. The key advantage
of the decomposition property is that it makes the optimization procedure highly efficient for
inverting the nonlinear function from the acoustic measurement to the VTR. For details, see a
recent publication in [110].
As an illustration, in Figs. 4.1–4.3, we plot the value of one term,
 
−π n fbs f
e cos 2π n ,
fs
in Eq. (4.46) as a function of the resonance frequency f and bandwidth b, for the first-order
(n = 1), second-order (n = 2), and the fifth-order (n = 5) cepstrum, respectively. (The sam-
pling frequency f s = 8000 Hz is used in all the plots.) These are the cepstra corresponding to
the transfer function of a single-resonance (i.e., one pole with no zeros) linear system. Due to
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

54 DYNAMIC SPEECH MODELS

Cepstral value for single resonance

Re
son
anc
eb z)
and cy (H
wid
th ( frequen
Hz n ance
) Reso

FIGURE 4.1: First-order cepstral value of a one-pole (single-resonance) filter as a function of the
resonance frequency and bandwidth. This plots the value of one term in Eq. (4.46) vs. f p and b p with
fixed n = 1 and f s = 8000 Hz

the decomposition property of the linear cepstrum, for multiple-resonance systems, the corre-
sponding cepstrum is simply a sum of those for the single-resonance systems.
Examining Figs. 4.1–4.3, we easily observe some key properties of the (single-resonance)
cepstrum. First, the mapping function from the VTR frequency and bandwidth variables to the
cepstrum, while nonlinear, is well behaved. That is, the relationship is smooth, and there is no
sharp discontinuity. Second, for a fixed resonance bandwidth, the frequency of the sinusoidal
relation between the cepstrum and the resonance frequency increases as the cepstral order
increases. The implication is that when piecewise linear functions are to be used to approximate
the nonlinear function of Eq. (4.46), more “pieces” will be needed for the higher-order than
for the lower-order cepstra. Third, for a fixed resonance frequency, the dependence of the low-
order cepstral values on the resonance bandwidth is relatively weak. The cause of this weak
dependence is the low ratio of the bandwidth (up to 800 Hz) to the sampling frequency (e.g.,
16 000 Hz) in the exponent of the cepstral expression in Eq. (4.46). For example, as shown
in Fig. 4.1 for the first-order cepstrum, the extreme values of bandwidths from 20 to 800 Hz
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

Cepstral value for single resonance

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 55

Re
son
anc
eb z)
and cy (H
wid
th ( frequen
Hz n ance
) Reso

FIGURE 4.2: Second-order cepstral value of a one-pole (single-resonance) filter as a function of the
resonance frequency and bandwidth (n = 1 and f s = 8000 Hz)

reduce the peak cepstral values only from 1.9844 to 1.4608 (computed by 2 exp(−20π/8000) and
2 exp(−800π/8000), respectively). The corresponding reduction for the second-order cepstrum
is from 0.9844 to 0.5335 (computed by exp(−2 × 20π/8000) and exp(−2 × 800π/8000),
respectively). In general, the exponential decay of the cepstral value, as the resonance bandwidth
increases, becomes only slightly more rapid for the higher-order than for the lower-order cepstra
(see Fig. 4.3). This weak dependence is desirable since the VTR bandwidths are known to
be highly variable with respect to the acoustic environment [120], and to be less correlated
with the phonetic content of speech and with human speech perception than are the VTR
frequencies.

Quantization Scheme for the Hidden Dynamic Vector

In the discretized hidden dynamic model, which is the theme of this chapter, the discretization
scheme is a central issue. We address this issue here using the example of the nonlinear function
discussed above, based on the recent work published in [110]. In that work, four poles are used
in the LPC model of speech [i.e., using P = 4 in Eq. (4.46)], since these lowest VTRs carry the
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

56 DYNAMIC SPEECH MODELS

Cepstral value for single resonance

Re
son
anc (Hz)
eb ency
and equ
wid
th ( nance fr
Hz
) Reso

FIGURE 4.3: Fifth-order cepstral value of a one-pole (single-resonance) filter as a function of the
resonance frequency and bandwidth n = 5 and f s = 8000 Hz

most important phonetic information of the speech signal. That is, an eight-dimensional vector
x = ( f 1 , f 2 , f 3 , f 4 , b 1 , b 2 , b 3 , b 4 ) is used as the input to the nonlinear function F(x). For the
output of the nonlinear function, up to 15 orders of linear cepstra are used. The zeroth order
cepstrum, c 0 , is excluded from the output vector, making the nonlinear mapping from VTRs
to cepstra independent of the energy level in the speech signal. This corresponds to setting the
gain G = 1 in the all-pole model of Eq. (4.42).
For each of the eight dimensions in the VTR vector, scalar quantization is used. Since
F(x) is relevant to all possible phones in speech, the appropriate range is chosen for each VTR
frequency and its corresponding bandwidth to cover all phones according to the considerations
discussed in [9]. Table 4.1 lists the range, from minimal to maximal frequencies in Hz, for
each of the four VTR frequencies and bandwidths. It also lists the corresponding number of
quantization levels used. Bandwidths are quantized uniformly with five levels while frequencies
are mapped to the Mel-frequency scale and then uniformly quantized with 20 levels. The
total number of quantization levels shown in Table 4.1 yields a total of 100 million (204 × 54 )
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 57

TABLE 4.1: Quantization Scheme for the VTR Variables, Including the Ranges of the Four VTR
Frequencies and Bandwidths and the Corresponding Numbers of Quantization Levels

MINIMUM (Hz) MAXIMUM (Hz) NO. OF QUANTIZATION

f1 200 900 20
f2 600 2800 20
f3 1400 3800 20
f4 1700 5000 20
b1 40 300 5
b2 60 300 5
b3 60 500 5
b4 100 700 5

entries for F(x), but because of the constraint f 1 < f 2 < f 3 < f 4 , the resulting number has
been reduced by about 25%.

4.2.4 E-Step for Parameter Estimation

After giving a comprehensive example above for the construction of a vector-valued nonlinear
mapping function and the quantization scheme for the vector valued hidden dynamics as the
input, we now return to the problem of parameter learning for the extended model. We also
return to the scalar case for the purpose of simplicity in exposition. We first describe the E-step
in the EM algorithm for the extended model, and concentrate on the differences from the basic
model as presented in a greater detail in the preceding section.
Like the basic model, before discretization, the auxiliary function for the E-step can be
simplified into the same form of
Q(r s , Ts , Bs , h s , Ds ) = Q x (r s , Ts , Bs ) + Q o (h s , Ds ) + Const., (4.47)
where

S 
N 
C 
C 
C
Q x (r s , Ts , Bs ) = 0.5 ξt (s , i, j, k) log |Bs |
s =1 t=1 i=1 j =1 k=1
2
−Bs (xt [i] − 2r s xt−1 [ j ] + r s2 xt−2 [k] − (1 − r s )2 Ts , (4.48)

and

S 
N 
C
Q o (h s , Ds ) = 0.5 γt (s , i) log |Ds | − Ds (o t − F(xt [i]) − h s )2 . (4.49)
s =1 t=1 i=1
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

58 DYNAMIC SPEECH MODELS

Again, large computational saving can be achieved by limiting the summations in Eq. (4.48)
for i, j, k based on the relative smoothness of trajectories in xt . That is, the range of i, j, k can
be set such that |xt [i] − xt−1 [ j ]| < Th1 , and |xt−1 [ j ] − xt−2 [k]| < Th2 . Now two thresholds,
instead of one in the basic model, are to be set.
In the above, we used ξt (s , i, j, k) and γt (s , i) to denote the frame-level posteriors of

ξt (s , i, j, k) ≡ p(s t = s , xt [i], xt−1 [ j ], xt−2 [k] | o 1N ),

and

γt (s , i) ≡ p(s t = s , xt [i] | o 1N ).

Note that ξt (s , i, j, k) has one more index k than the counterpart in the basic model. This is
due to the additional conditioning in the second-order state equation.
Similar to the basic model, in order to compute ξt (s , i, j, k) and γt (s , i), we need to
compute the forward and backward probabilities by recursion. The forward recursion αt (s , i) ≡
p(o 1t , s t = s , it = i) is


S 
C
α(s t+1 , it+1 ) = α(s t , it ) p(s t+1 , it+1 | s t , it , it−1 ) p(o t+1 | s t+1 , it+1 ), (4.50)
s t =1 it =1

where

p(o t+1 | s t+1 = s , it+1 = i) = N(o t+1 ; F(xt+1 [i]) + h s , Ds ),

and

p(s t+1 = s , it+1 = i | s t = s  , it = j, it−1 = k)

≈ p(s t+1 = s | s t = s  ) p(it+1 = i | it = j, it−1 = k)
= πs  s N(xt [i]; 2r s xt−1 [ j ] − r s2 xt−2 [k] + (1 − r s )2 Ts , Bs ).

The backward recursion βt (s , i) ≡ p(o t+1

N
| s t = s , it = i) is


S 
C
β(s t , it ) = β(s t+1 , it+1 ) p(s t+1 , it+1 | s t , it , it−1 ) p(o t+1 | s t+1 , it+1 ). (4.51)
s t+1 =1 it+1 =1

Given αt (s , i) and β(s t , it ) as computed, we can obtain the posteriors of ξt (s , i, j, k) and

γt (s , i).
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 59

4.2.5 M-Step for Parameter Estimation
Reestimation for Parameter rs
To obtain the reestimation formula for parameter r s , we set the following partial derivative to
zero:

∂ Q x (r s , Ts , Bs ) N  C  C  C
= −Bs ξt (s , i, j, k) (4.52)
∂r s t=1 i=1 j =1 k=1

× xt [i] − 2r s xt−1 [ j ] + r s2 xt−2 [k] − (1 − r s )2 Ts −xt−1 [ j ] + r s xt−2 [k] + (1 − r s )Ts


N 
C 
C 
C
= −Bs ξt (s , i, j, k)
t=1 i=1 j =1 k=1

× −xt [i]xt−1 [ j ] + 2r s xt−1

2
[ j ] − r s2 xt−1 [ j ]xt−2 [k] + (1 − r s )2 xt−1 [ j ]Ts
+ r s xt [i]xt−2 [k] − 2r s2 xt−1 [ j ]xt−2 [k] + r s3 xt−2
2
[k] − r s (1 − r s )2 xt−2 [k]Ts
+ xt [i](1 − r s )Ts − 2r s xt−1 [ j ](1 − r s )Ts + r s2 xt−2 [k](1 − r s )Ts − (1 − r s )3 T 2s = 0.

This can be written in the following form in order to solve for r s (assuming Ts is fixed
from the previous EM iteration):

A3r̂ s3 + A2r̂ s2 + A1r̂ s + A0 = 0, (4.53)

where

N 
C 
C 
C
A3 = ξt (s , i, j, k){xt−2
2
[k] + Ts xt−2 [k] + Ts 2 },
t=1 i=1 j =1 k=1


N 
C 
C 
C
A2 = ξt (s , i, j, k){−3xt−1 [ j ]xt−2 [k] + 3Ts xt−1 [ j ] + 3Ts xt−2 [k] − 3Ts 2 },
t=1 i=1 j =1 k=1


N 
C 
C 
C
A1 = ξt (s , i, j, k){2xt−1
2
[ j ] + xt [i]xt−2 [k] − xt [i]Ts
t=1 i=1 j =1 k=1

− 4xt−1 [ j ]Ts − xt−2 [k]Ts + 3Ts 2 },


N 
C 
C 
C
A0 = ξt (s , i, j, k){−xt [i]xt−1 [ j ] + xt [i]Ts + xt−1 [ j ]Ts − Ts 2 }. (4.54)
t=1 i=1 j =1 k=1

Analytic solutions exist for third-order algebraic equations such as the above. For the three roots
found, constraints 1 > r s > 0 can be used for selecting the appropriate one. If there is more
than one solution satisfying the constraint, then we can select the one that gives the largest
value for Q x .
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

60 DYNAMIC SPEECH MODELS

Reestimation for Parameter Ts
We now optimize Ts by setting the following partial derivative to zero:

∂ Q x (r s , Ts , Bs ) N  C  C  C
= −Bs ξt (s , i, j, k)[xt [i]
∂ Ts t=1 i=1 j =1 k=1
−2r s xt−1 [ j ] + r s2 xt−2 [k] − (1 − r s )2 Ts ](1 − r s )2 = 0. (4.55)

Now fixing r s from the previous EM iteration, we obtain an explicit solution to the reestimate
of Ts :

1 N  C  C  C
T̂s = ξt (s , i, j, k){xt [i] − 2r s xt−1 [ j ] + r s2 xt−2 [k]}.
(1 − r s )2 t=1 i=1 j =1 k=1

Reestimation for Parameter h s

We set

∂ Q o (h s , Ds ) N  C
= −Ds γt (s , i){o t − F(xt [i]) − h s } = 0. (4.56)
∂h s t=1 i=1

This gives the reestimation formula:

 N C
i=1 γt (s , i){o t − F(xt [i])}
ĥ s = t=1
 N C . (4.57)
t=1 i=1 γt (s , i)

Reestimation for Bs and Ds

Setting

∂ Q x (r s , Ts , Bs ) N  C  C  C
= 0.5 ξt (s , i, j, k)[Bs−1
∂ Bs t=1 i=1 j =1 k=1
2
− xt [i] − 2r s xt−1 [ j ] + r s2 xt−2 [k] − (1 − r s )2 Ts ] = 0, (4.58)

we obtain the reestimation formula:

 N C C C  2
k=1 ξt (s , i, j, k) xt [i] − 2r s xt−1 [ j ] + r s xt−2 [k] − (1 − r s ) Ts
2 2
t=1 i=1 j =1
B̂s =  N C C C .
t=1 i=1 j =1 k=1 ξ t (s , i, j, k)
(4.59)
Similarly, setting

∂ Q o (Hs , h s , Ds ) N  C
= 0.5 γt (s , i) Ds−1 − (o t − Hs xt [i] − h s )2 = 0,
∂ Ds t=1 i=1
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 61

we obtain the reestimate (scalar value) of
 N C
γt (s , i) [o t − Hs xt [i] − h s ]2
D̂s = t=1 i=1 N C . (4.60)
t=1 i=1 γt (s , i)

4.2.6 Decoding of Discrete States by Dynamic Programming

The DP recursion is essentially the same as in the basic model, except an additional level (index
k) of optimization is introduced due to the second-order dependency in the state equation. The
final form of the recursion can be written as

δt+1 (s , i) = max
 
δt (s  , i  ) p(s t+1 = s , it+1 = i | s t = s  , it = i  ) p(o t+1 | s t+1 = s , it+1 = i)
s ,i
≈ max
 
δt (s  , i  ) p(s t+1 = s | s t = s  ) p(it+1 = i | it = i  ) p(o t+1 | s t+1 = s , it+1 = i)
s ,i

= max

δt (s  , i  )πs  s N(xt+1 [i]; 2r s  xt [ j ] − r s2 xt−1 [k] + (1 − r s  )2 Ts , Bs )
s ,i , j,k
×N(o t+1 ; F(xt+1 [i]) + h s , Ds ). (4.61)

4.3 APPLICATION TO AUTOMATIC TRACKING

OF HIDDEN DYNAMICS
As an example for the application of the discretized hidden dynamic model discussed in this
chapter so far, we discuss implementation efficiency issues and show results for the specific
problem of automatic tracking of the hidden dynamic variables that are discretized. The accuracy
of the tracking is obviously limited by the discretization level, but this approximation makes it
possible to run the parameter learning and decoding algorithms in a manner that is not only
tractable but also efficient.
While the description of the parameter learning and decoding algorithms earlier in
this chapter is confined to the scalar case for purposes of clarification and notational con-
venience, in practical cases where often vector valued hidden dynamics are involved, we need
to address the problem of algorithms’ efficiency. In the application example in this section
where eight-dimensional hidden dynamic vectors (four VTR frequencies and four bandwidths
x = ( f 1 , f 2 , f 3 , f 4 , b 1 , b 2 , b 3 , b 4 )) are used as presented in detail in Section 4.2.3, it is important
to address the issue related to the algorithms’ efficiency.

4.3.1 Computation Efficiency: Exploiting Decomposability in the

Observation Function
For multidimensional hidden dynamics, one obvious difficulty for the training and tracking
algorithms presented earlier is the high computational cost in summing and in searching over
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

62 DYNAMIC SPEECH MODELS

the huge space in the quantized hidden dynamic variables. The sum with C terms as required
in the various reestimation formulas and in the dynamic programming recursion is typically
expensive since C is very large. With scalar quantization for each of the eight VTR dimensions,
the C would be the Cartesian product of the quantization levels for each of the dimensions.
To overcome this difficulty, a suboptimal, greedy technique is implemented as described
in [110]. This technique capitalizes on the decomposition property of the nonlinear mapping
function from VTR to cepstra that we described earlier in Section 4.2.3. This enables a much
smaller number of terms to be evaluated than the rigorous number determined as the Cartesian
product, which we elaborate below.
Let us consider an objective function F, to be optimized with respect to M noninteracting
or decomposable variables that determine the function’s value. An example is the following
decomposable function consisting of M terms Fm , m = 1, 2, . . . , M, each of which contains
independent variables (αm ) to be searched for:

M
F= Fm (αm ).
m=1

Note that the VTR-to-cepstrum mapping function, which was derived to be Eq. (4.46)
as the observation equation of the dynamic speech model (extended model), has this de-
composable form. The greedy optimization technique proceeds as follows. First, initialize
αm , m = 1, 2, . . . , M to reasonable values. Then, fix all αm s except one, say αn , and optimize
αn with respect to the new objective function of

n−1 
M
F− Fm (αm ) − Fm (αm ).
m=1 m=n+1

Next, after the low-dimensional, inexpensive search problem for αˆn is solved, fix it and
optimize a new αm , m = n. Repeat this for all αm s . Finally, iterate the above process until all
optimized αm s become stabilized.
In the implementation of this technique for VTR tracking and parameter estimation as
reported in [110], each of the P = 4 resonances is treated as a separate, noninteractive variables
to optimize. It was found that only two to three overall iterations above are already sufficient to
stabilize the parameter estimates. (During the training of the residual parameters, these inner
iterations are embedded in each of the outer EM iterations.) Also, it was found that initialization
of all VTR variables to zero gives virtually the same estimates as those by more carefully thought
initialization schemes.
With the use of the above greedy, suboptimal technique instead of full optimal search,
the computation cost of VTR tracking was reduced by over 4000-fold compared with the
brute-force implementation of the algorithms.
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 63

4.3.2 Experimental Results
As reported in [110], the above greedy technique was incorporated into the VTR tracking
algorithm and into the EM training algorithm for the nonlinear-prediction residual parameters.
The state equation was made simpler than the counterpart in the basic or extended model in
that all the phonological states s are tied. This is because for the purposes of tracking hidden
dynamics there is no need to distinguish the phonological states. The DP recursion in the more
general case of Eq. (4.33) can then be simplified by eliminating the optimization on index s ,
leaving only the indices i and j of the discretization levels in the hidden VTR variables during
the DP recursion. We also set the parameter r s = 1 uniformly in all the experiments. This gives
the role of the state equation as a “smoothness” constraint.
The effectiveness of the EM parameter estimation, Eqs. (4.57) and (4.60) in particular,
discussed for the extended model in this chapter will be demonstrated in the VTR tracking
experiments. Due to the tying of the phonological states, the training does not require any
data labeling and is fully unsupervised. Fig. 4.4 shows the VTR tracking ( f 1 , f 2 , f 3 , f 4 ) results,

FIGURE 4.4: VTR tracking by setting the residual mean vector to zero
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

64 DYNAMIC SPEECH MODELS

superimposed on the spectrogram of a telephone speech utterance (excised from the Switchboard
database) of “the way you dress” by a male speaker, when the residual mean vector h (tied over
all s state) was set to zero and the covariance matrix D is set to be diagonal with empirically
determined diagonal values. [The initialized variances are those computed from the codebook
entries that are constructed from quantizing the nonlinear function in Eq. (4.46.)] Setting h
to zero corresponds to the assumption that the nonlinear function of Eq. (4.46) is an unbiased
predictor of the real speech data in the form of linear cepstra. Under this assumption we observe
from Fig. 4.4 that while f 1 and f 2 are accurately tracked through the entire utterance, f 3 and
f 4 are incorrectly tracked during the later half of the utterance. (Note that the many small step
jumps in the VTR estimates are due to the quantization of the VTR frequencies.) One iteration
of the EM training on the residual mean vector and covariance matrix does not correct the
errors (see Fig. 4.5), but two iterations are able to correct the errors in the utterance for about
20 frames (after time mark of 0.6 s in Fig. 4.6). One further iteration is able to correct almost
all errors as shown in Fig. 4.7.

FIGURE 4.5: VTR tracking with one iteration of residual training

P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 65

FIGURE 4.6: VTR tracking with two iterations of residual training

To examine the quantitative behavior of the residual parameter training, we list the log-
likelihood score as a function of the EM iteration number in Table 4.2. Three iterations of the
training appear to have reached the EM convergence. When we examine the VTR tracking
results after 5 and 20 iterations, they are found to be identical to Fig. 4.7, consistent with the
near-constant converging log-likelihood score reached after three iterations of training. Note
that the regions in the utterance where the speech energy is relatively low are where consonantal
constriction or closure is formed; e.g., near time mark of 0.1 s for /w/ constriction and near
time mark of 0.4 s for /d/ closure). The VTR tracker gives almost as accurate estimates for the
resonance frequencies in these regions as for the vowel regions.

4.4 SUMMARY
This chapter discusses one of the two specific types of hidden dynamic models in this book,
as example implementations of the general modeling and computational scheme introduced in
Chapter 2. The essence of the implementation described in this chapter is the discretization
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

66 DYNAMIC SPEECH MODELS

FIGURE 4.7: VTR tracking with three iterations of residual training

TABLE 4.2: Log-likelihood Score as a Function of the EM

Iteration Number in Training the Nonlinear-prediction Resid-
ual Parameters

EM ITERATION LOG-LIKELIHOOD
NO. SCORE
0 1.7680
1 2.0813
2 2.0918
3 2.1209
5 2.1220
20 2.1222
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

MODELS WITH DISCRETE-VALUED HIDDEN SPEECH DYNAMICS 67

of the hidden dynamic variables. While this implementation introduces approximations to
the original continuous-valued variables, the otherwise intractable parameter estimation and
decoding algorithms have become tractable, as we have presented in detail in this chapter.
This chapter starts by introducing the “basic” model, where the state equation in the
dynamic speech model gives discretized first-order dynamics and the observation equation is a
linear relationship between the discretized hidden dynamic variables and the acoustic observa-
tion variables. Probabilistic formulation of the model is presented first, which is equivalent to
the state–space formulation but is in a form that can be more readily used in developing and
describing the model parameter estimation algorithms. The parameter estimation algorithms
are presented, with sufficient detail in deriving all the final reestimation formulas as well as the
key intermediate quantities such as the auxiliary function in the E-step of the EM algorithm.
In particular, we separate the forward–backward algorithm out of the general E-step derivation
in a new subsection to emphasize its critical role. After deriving the reestimation formulas for
all model parameters as the M-step, we describe a DP-based algorithm for jointly decoding
the discrete phonological states and the hidden dynamic “state,” the latter constructed from
discretization of the continuous variables.
The chapter is followed by presenting an extension of the basic model in two aspects.
First, the state equation is extended from the first-order dynamics to the second-order dy-
namics, making the shape of the temporally unfolded “trajectories” more realistic. Second, the
observation equation is extended from the linear mapping to a nonlinear one. A new subsection
is then devoted to a special construction of the nonlinear mapping where a “physically” based
prediction function is developed when the hidden dynamic variables as the input are taken to
be the VTRs and the acoustic observations as the output are taken to be the linear cepstral
features. Using this nonlinear mapping function, we proceed to develop the E-step and M-step
of the EM algorithm for this extended model in a way parallel to that for the basic model.
Finally, we give an application example of the use of a simplified version of the extended
model and the related algorithms discussed in this chapter for automatic tracking of the hidden
dynamic vectors, the VTR trajectories in this case. Specific issues related to the tracking algo-
rithm’s efficiency arising from multidimensionality in the hidden dynamics are addressed, and
experimental results on some typical outputs of the algorithms are presented and analyzed.
P1: IML/FFX P2: IML
MOBK024-04 MOBK024-LiDeng.cls May 30, 2006 15:30

68
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

69

CHAPTER 5

Models with Continuous-Valued

Hidden Speech Trajectories

The preceding chapter discussed the implementation strategy for hidden dynamic models based
on discretizing the hidden dynamic values. This permits tractable but approximate learning of
the model parameters and decoding of the discrete hidden states (both phonological states
and discretized hidden dynamic “states”). This chapter elaborates on another implementation
strategy where the continuous-valued hidden dynamics remain unchanged but a different type of
approximation is used. This implementation strategy assumes fixed discrete-state (phonological
unit) boundaries, which may be obtained initially from a simpler speech model set such as the
HMMs and then be further refined after the dynamic model is learned iteratively. We will
describe this new implementation and approximation strategy for a hidden trajectory model
(HTM) where the hidden dynamics are defined as an explicit function of time instead of by
recursion. Other types of approximation developed for the recursively defined dynamics can be
found in [84, 85, 121–123] and will not be described in this book.
This chapter extracts, reorganizes, and expands the materials published in [109,115,116,
124], fitting these materials into the general theme of dynamic speech modeling in this book.

5.1 OVERVIEW OF THE HIDDEN TRAJECTORY MODEL

As a special type of the hidden dynamic model, the HTM presented in this section is a struc-
tured generative model, from the top level of phonetic specification to the bottom level of
acoustic observations via the intermediate level of (nonrecursive) FIR-based target filtering that
generates hidden VTR trajectories. One advantage of the FIR filtering is its natural handling of
the two constraints (segment-bound monotonicity and target-directedness) that often requires
asynchronous segment boundaries for the VTR dynamics and for the acoustic observations.
This section is devoted to the mathematical formulation of the HTM as a statistical
generative model. Parameterization of the model is detailed here, with consistent notations set
up to facilitate the derivation and description of algorithmic learning of the model parameters
presented in the next section.
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

70 DYNAMIC SPEECH MODELS

5.1.1 Generating Stochastic Hidden Vocal Tract Resonance Trajectories
The HTM assumes that each phonetic unit is associated with a multivariate distribution of
the VTR targets. (There are exceptions for several compound phonetic units, including diph-
thongs and affricates, where two distributions are used.) Each phone-dependent target vector, ts ,
consists of four low-order resonance frequencies appended by their corresponding bandwidths,
where s denotes the segmental phone unit. The target vector is a random vector—hence stochas-
tic target—whose distribution is assumed to be a (gender-dependent) Gaussian:
p(t | s ) = N (t; μTs , ΣTs ). (5.1)
The generative process in the HTM starts by temporal filtering the stochastic targets.
This results in a time-varying pattern of stochastic hidden VTR vectors z (k). The filter is
constrained so that the smooth temporal function of z (k) moves segment-by-segment towards
the respective target vector t s but it may or may not reach the target depending on the degree
of phonetic reduction.
These phonetic targets are segmental in that they do not change over the phone segment
once the sample is taken, and they are assumed to be largely context-independent. In our HTM
implementation, the generation of the VTR trajectories from the segmental targets is through
a bidirectional finite impulse response (FIR) filtering. The impulse response of this noncausal
filter is
⎧
−k
⎪
⎨ c γ s (k) −D < k < 0,
h s (k) = c k = 0, (5.2)
⎪
⎩ c γk
s (k) 0 < k < D,
where k represents time frame (typically with a length of 10 ms each), and γs (k) is the segment-
dependent “stiffness” parameter vector, one component for each resonance. Each component
is positive and real-valued, ranging between zero and one. In Eq. (5.2), c is a normalization
constant, ensuring that h s (k) sums to one over all time frames k. The subscript s (k) in γs (k)
indicates that the stiffness parameter is dependent on the segment state s (k), which varies
over time. D in Eq. (5.2) is the unidirectional length of the impulse response, representing the
temporal extent of coarticulation in one temporal direction, assumed for simplicity to be equal
in length for the forward direction (anticipatory coarticulation) and the backward direction
(regressive coarticulation).
In Eq. (5.2), c is the normalization constant to ensure that the filter weights add up to
one. This is essential for the model to produce target undershooting, instead of overshooting.
To determine c , we require that the filter coefficients sum to one:

D 
D
|k|
h s (k) = c γ s (k) = 1. (5.3)
k=−D k=−D
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 71

For simplicity, we make the assumption that over the temporal span of −D ≤ k ≤ D, the
stiffness parameter’s value stays approximately constant
γ s (k) ≈ γ.
That is, the adjacent segments within the temporal span of 2D + 1 in length that contribute
to the coarticulated home segment have the same stiffness parameter value as that of the home
segment. Under this assumption, we simplify Eq. (5.3) to

D
|k| 1 + γ − 2γ D+1
c γ s (k) ≈ c [1 + 2(γ + γ 2 + · · · + γ D )] = c .
k=−D
1−γ

Thus,
1−γ
c (γ) ≈ . (5.4)
1 + γ − 2γ D+1
The input to the above FIR filter as a linear system is the target sequence, which is a
function of discrete time and is subject to abrupt jumps at the phone segments’ boundaries.
Mathematically, the input is represented as a sequence of stepwise constant functions with
variable durations and heights:

I
t(k) = [u(k − ksl i ) − u(k − ksri )]t s i , (5.5)
i=1

where u(k) is the unit step function, k sr , s = s 1 , s 2 , . . . , s I are the right boundary sequence
of the segments (I in total) in the utterance, and ksl , s = s 1 , s 2 , . . . , s I are the left boundary
sequence. Note the constraint on these starting and end times: k sl +1 = k sr . The difference of
the two boundary sequences gives the duration sequence. t s , s = s 1 , s 2 , . . . , s I are the random
target vectors for segment s .
Given the filter’s impulse response and the input to the filter as the segmental VTR
target sequence t(k), the filter’s output as the model’s prediction for the VTR trajectories is the
convolution between these two signals. The result of the convolution within the boundaries of
home segment s is

k+D
|k−τ |
z (k) = h s (k) ∗ t(k) = c γ γ s (τ ) t s (τ ) , (5.6)
τ =k−D

where the input target vector’s value and the filter’s stiffness vector’s value typically take not only
those associated with the current home segment, but also those associated with the adjacent
segments. The latter case happens when the time τ in Eq. (5.6) goes beyond the home segment’s
boundaries, i.e., when the segment s (τ ) occupied at time τ switches from the home segment to
an adjacent one.
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

72 DYNAMIC SPEECH MODELS

The linearity between z and t as in Eq. (5.6) and Gaussianity of the target t makes the
VTR vector z (k) (at each frame k) a Gaussian as well. We now discuss the parameterization of
this Gaussian trajectory:

p(z (k) | s ) = N [z (k); μz(k) , Σz(k) ]. (5.7)

The mean vector above is determined by the filtering function:


k+D
|k−τ |
μz(k) = c γ γ s (τ ) μTs (τ ) = a k · μT . (5.8)
τ =k−D

Each f th component of vector μz (k) is


L
μz (k) ( f ) = a k (l)μT (l, f ), (5.9)
l=1

where L is the total number of phone-like HTM units as indexed by l, and f = 1, . . . , 8 denotes
four VTR frequencies and four corresponding bandwidths.
The covariance matrix in Eq. (5.7) can be similarly derived to be

k+D
2|k−τ |
Σz(k) = c γ2 γs (τ ) ΣTs (τ ) .
τ =k−D

Approximating the covariance matrix by a diagonal one for each phone unit l, we represent its
diagonal elements as a vector:

σ 2z(k) = vk · σ 2T . (5.10)

and the target covariance matrix is also approximated as diagonal:

⎡ ⎤
σT2 (l, 1) 0 ··· 0
⎢ ⎥
⎢ 0 σT2 (l, 2) · · · 0 ⎥
ΣT (l) ≈ ⎢⎢ ⎥.
.. .. .. .. ⎥
⎣ . . . . ⎦
0 0 · · · σT2 (l, 8)

The f th element of the vector in Eq. (5.10) is


L
σz2(k) ( f) = vk (l)σT2 (l, f ). (5.11)
l=1

In Eqs. (5.8) and (5.10), a k and vk are frame (k)-dependent vectors. They are constructed
for any given phone sequence and phone boundaries within the coarticulation range (2D + 1
frames) centered at frame k. Any phone unit beyond the 2D + 1 window contributes a zero
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 73

value to these “coarticulation” vectors’ elements. Both a k and vk are a function of the phones’
identities and temporal orders in the utterance, and are independent of the VTR dimension f .

5.1.2 Generating Acoustic Observation Data

The next generative process in the HTM provides a forward probabilistic mapping or prediction
from the stochastic VTR trajectory z (k) to the stochastic observation trajectory o(k). The
observation takes the form of linear cepstra. An analytical form of the nonlinear prediction
function F[z (k)] presented here is in the same form as described (and derived) in Section 4.2.3
of Chapter 4 and is summarized here:

2 P b p (k)
−πq fsamp f p (k)
Fq (k) = e cos(2πq ), (5.12)
q p=1 f samp

where f samp is the sampling frequency, P is the highest VTR order (P = 4), and q is the cepstral
order.
We now introduce the cepstral prediction’s residual vector:

rs (k) = o(k) − F[z (k)].

We model this residual vector as a Gaussian parameterized by residual mean vector μr s (k) and
covariance matrix Σr s (k) :

p(rs (k) | z (k), s ) = N rs (k); μr s (k) , Σr s (k) . (5.13)

Then the conditional distribution of the observation becomes:


p(o(k) | z (k), s ) = N o(k); F[z (k)] + μr s (k) , Σr s (k) . (5.14)

An alternative form of the distribution in Eq. (5.14) is the following “observation equa-
tion”:

o(k) = F[z (k)] + μr s (k) + vs (k),

where the observation noise vs (k) ∼ N (vs ; 0, Σr s (k) ).

5.1.3 Linearizing Cepstral Prediction Function

To facilitate computing the acoustic observation (linear cepstra) likelihood, it is important to
characterize the linear cepstra uncertainty in terms of its conditional distribution on the VTR,
and to simplify the distribution to a computationally tractable form. That is, we need to specify
and approximate p(o | z, s ). We take the simplest approach to linearize the nonlinear mean
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

74 DYNAMIC SPEECH MODELS

function of F[z (k)] in Eq. (5.14) by using the first-order Taylor series approximation:

F[z (k)] ≈ F[z0 (k)] + F  [z0 (k)](z (k) − z0 (k)), (5.15)

where the components of Jacobian matrix F  [·] can be computed in a closed form of
 
4π −πq bfsamp
p (k) f p (k)
Fq [ f p (k)] =− e sin 2πq , (5.16)
f samp f samp

for the VTR frequency components of z , and

 
2π −πq bfsamp
p (k) f p (k)
Fq [b p (k)] =− e cos 2πq , (5.17)
f samp f samp

for the VTR bandwidth components of z . In the current implementation, the Taylor series
expansion point z0 (k) in Eq. (5.15) is taken as the tracked VTR values based on the HTM.
Substituting Eq. (5.15) into Eq. (5.14), we obtain the approximate conditional acoustic
observation probability where the mean vector μo s is expressed as a linear function of the VTR
vector z :

p(o(k) | z (k), s ) ≈ N (o(k); μo s (k) , Σr s (k) ), (5.18)

where

μo s (k) = F  [z 0 (k)]z (k) + F[z 0 (k)] − F  [z 0 (k)]z 0 (k) + μr s (k) . (5.19)

This then permits a closed-form solution for acoustic likelihood computation, which we
derive now.

5.1.4 Computing Acoustic Likelihood

An essential aspect of the HTM is its ability to provide the likelihood value for any sequence of
acoustic observation vectors o(k) in the form of cepstral parameters. The efficiently computed
likelihood provides a natural scoring mechanism comparing different linguistic hypotheses as
needed in speech recognition. No VTR values z (k) are needed in this computation as they are
treated as the hidden variables. They are marginalized (i.e., integrated over) in the linear cepstra
likelihood computation. Given the model construction and the approximation described in the
preceding section, the HTM likelihood computation by marginalization can be carried out in
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 75

a closed form. Some detailed steps of derivation give

p(o(k) | s ) = p[o(k) | z (k), s ] p[z (k) | s ] d z

≈ N [o(k); μo s (k) , Σr s (k) ] N [z (k); μz(k) , Σz(k) ] d z
 
= N o(k); μ̄o s (k) , Σ̄o s (k) , (5.20)

where the time (k)-varying mean vector is

μ̄o s (k) = F[z0 (k)] + F  [z0 (k)][a k · μT − z0 (k)] + μr s (k) , (5.21)

and the time-varying covariance matrix is

Σ̄o s (k) = Σr s (k) + F  [z0 (k)]Σz(k)(F  [z0 (k)])Tr . (5.22)

The final result of Eqs. (5.20)–(5.22) are quite intuitive. For instance, when the Taylor
series expansion point is set at z0 (k) = μz(k) = a k · μT , Eq. (5.21) is simplified to μ̄o s (k) =
F[μz(k)] + μr s , which is the noise-free part of cepstral prediction. Also, the covariance ma-
trix in Eq. (5.20) is increased by the quantity F  [z0 (k)]Σz(k)(F  [z0 (k)])Tr over the covariance
matrix for the cepstral residual term Σr s (k) only. This magnitude of increase reflects the newly
introduced uncertainty in the hidden variable, measured by Σz(k). The variance amplification
factor F  [z0 (k)] results from the local “slope” in the nonlinear function F[z] that maps from
the VTR vector z (k) to cepstral vector o(k).
It is also interesting to interpret the likelihood score Eq. (5.20) as probabilistic charac-
terization of a temporally varying Gaussian process, where the time-varying mean vectors are
expressed in Eq. (5.21) and the time-varying covariance matrices are expressed in Eq. (5.22).
This may make the HTM look ostensibly like a nonstationary-state HMM (within the acoustic
dynamic model category). However, the key difference is that in HTM the dynamic structure
represented by the hidden VTR trajectory enters into the time-varying mean vector Eq. (5.21)
in two ways: (1) as the argument z0 (k) in the nonlinear function F[z0 (k)]; and (2) as the
term a k · μT = μz(k) in Eq. (5.21). Being closely related to the VTR tracks, they both capture
long-span contextual dependency, yet with mere context-independent VTR target parameters.
Similar properties apply to the time-varying covariance matrices in Eq. (5.22). In contrast, the
time-varying acoustic dynamic models do not have these desirable properties. For example, the
polynomial trajectory model [55, 56, 86] does regression fitting directly on the cepstral data,
exploiting no underlying speech structure and hence requiring context dependent polynomial
coefficients for representing coarticulation. Likewise, the more recent trajectory model [26] also
relies on a very large number of free model parameters to capture acoustic feature variations.
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

76 DYNAMIC SPEECH MODELS

5.2 UNDERSTANDING MODEL BEHAVIOR

BY COMPUTER SIMULATION
In this section, we present the model simulation results, extracted from the work published
in [109], demonstrating major dynamic properties of the HTM. We further compare these
results with the corresponding results from direct measurements of reduction in the acoustic–
phonetic literature.
To illustrate VTR frequency or formant target undershooting, we first show the spectro-
gram of three renditions of a three-segment /iy aa iy/ (uttered by the author of this book) in
Fig. 5.1. From left to right, the speaking rate increases and speaking effort decreases, with the
durations of the /aa/’s decreasing from approximately 230 to 130 ms. Formant target under-
shooting for f 1 and f 2 is clearly visible in the spectrogram, where automatically tracked formants
are superimposed (as the solid lines) in Fig. 5.1 to aid identification of the formant trajectories.
(The dashed lines are the initial estimates, which are then refined to give the solid lines.)

5.2.1 Effects of Stiffness Parameter on Reduction

The same kind of target undershooting for f 1 and f 2 as in Fig. 5.1 is exhibited in the model
prediction, shown in Fig. 5.2, where we also illustrate the effects of the FIR filter’s stiffness
parameter on the magnitude of formant undershooting or reduction. The model prediction
is the FIR filter’s output for f 1 and f 2 . Figs. 5.2(a)–(c) correspond to the use of the stiffness
parameter value (the same for each formant vector component) set at γ = 0.85, 0.75 and 0.65,
respectively, where in each plot the slower /iy aa iy/ sounds (with the duration of /aa/ set at

FIGURE 5.1: Spectrogram of three renditions of /iy aa iy/ by one author, with an increasingly higher
speaking rate and increasingly lower speaking efforts. The horizontal label is time, and the vertical one
is frequency
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 77

(a) γ = [0.85], D=100
2500
/iy/
2000 /iy/ /iy/
f2 (Hz)
/a/ /a/
1500
1000
500 f1 (Hz)

(b) γ = [0.75]
2500
2000
1500
1000
500
0

(c) γ = [0.65]
2500
2000
1500
1000
500
0
0 20 40 60 80 100 120
Time frame (0.01 s)

FIGURE 5.2: f 1 and f 2 formant or VTR frequency trajectories produced from the model for a slow /iy
aa iy/ followed by a fast /iy aa iy/. (a), (b) and (c) correspond to the use of the stiffness parameter values
of γ = 0.85, 0.75 and 0.65, respectively. The amount of formant undershooting or reduction during the
fast /aa/ is decreasing as the γ value decreases. The dashed lines indicate the formant target values and
their switch at the segment boundaries

230 ms or 23 frames) are followed by the faster /iy aa iy/ sounds (with the duration of /aa/ set
at 130 ms or 13 frames). f 1 and f 2 targets for /iy/ and /aa/ are set appropriately in the model
also. Comparing the three plots, we have the model’s quantitative prediction for the magnitude
of reduction in the faster /aa/ that is decreasing as the γ value decreases.
In Figs. 5.3(a)–(c), we show the same model prediction as in Fig. 5.2 but for different
sounds /iy eh iy/, where the targets for /eh/ are much closer to those of the adjacent sound /iy/
than in the previous case for /aa/. As such, the absolute amount of reduction becomes smaller.
However, the same effect of the filter parameter’s value on the size of reduction is shown as for
the previous sounds /iy aa iy/.
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

78 DYNAMIC SPEECH MODELS

(a) γ = [0.85], D=100
2500
/ε/ /ε/
2000 /iy/ /iy/ /iy/
1500
1000
500
0

(b) γ = [0.75]
2500
2000
1500
1000
500
0

(c) γ = [0.65]
2500
2000
1500
1000
500
0
0 20 40 60 80 100 120
Time frame

FIGURE 5.3: Same as Fig. 5.2 except for the /iy eh iy/ sounds. Note that the f 1 and f 2 target values
for /eh/ are closer to /iy/ than those for /aa/

5.2.2 Effects of Speaking Rate on Reduction

In Fig. 5.4, we show the effects of speaking rate, measured as the inverse of the sound segment’s
duration, on the magnitude of formant undershooting. Subplots (a)–(c) correspond to three
decreasing durations of the sound /aa/ in the /iy aa iy/ sound sequence. They illustrate an
increasing amount of the reduction with the decreasing duration or increasing speaking rate.
Symbol “x” in Fig. 5.4 indicates the f 1 and f 2 formant values at the central portions of vowels/
aa/, which are predicted from the model and are used to quantify the magnitude of reduction.
These values (separately for f 1 and f 2 ) for /aa/ are plotted against the inversed duration in
Fig. 5.5, together with the corresponding values for /eh/ (i.e., IPA ) in the /iy eh iy/ sound
sequence. The most interesting observation is that as the speaking rate increases, the distinction
between vowels /aa/ and /eh/ gradually diminishes if their static formant values extracted from
the dynamic patterns are used as the sole measure for the difference between the sounds. We
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 79

(a) γ = [0.85], D = 100
2500
2000
1500 x

1000
x
500
0

(b) γ = [0.85]
2500
2000
x
1500
1000
x
500
0

(c) γ = [0.85]
2500
2000
x
1500
1000
x
500
0
0 10 20 30 40 50 60

FIGURE 5.4: f 1 and f 2 formant trajectories produced from the model for three different durations of
/aa/ in the /iy aa iy/ sounds: (a) 25 frames (250 ms), (b) 20 frames and (c) 15 frames. The same γ value of
0.85 is used. The amount of target undershooting increases as the duration is shortened or the speaking
rate is increased. Symbol “x” indicates the f 1 and f 2 formant values at the central portions of vowels of /aa/

refer to this phenomenon as “static” sound confusion induced by increased speaking rate (or/and
by a greater degree of sloppiness in speaking).

5.2.3 Comparisons with Formant Measurement Data

The “static” sound confusion between /aa/ and /eh/ quantitatively predicted by the model
as shown in Fig. 5.5 is consistent with the formant measurement data published in [125],
where thousands of natural sound tokens were used to investigate the relationship between
the degree of formant undershooting and speaking rate. We reorganized and replotted the raw
data from [125] in Fig. 5.6, in the same formant as Fig. 5.5. While the measures of speaking
rate differ between the measurement data and model prediction and cannot be easily converted
to each other, they are generally consistent with each other. The similar trend for the greater
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

80 DYNAMIC SPEECH MODELS

2000

1800 /ε/

1600
f2
/a/
1400
Predicted formant frequencies (Hz)

1200

1000

800
/a/

600 f1

400 /ε/

200
2 3 4 5 6 7 8 9 10

Speaking rate (inverse of duration in s)

FIGURE 5.5: Relationship, based on model prediction, between the f 1 and f 2 formant values at the
central portions of vowels and the speaking rate. Vowel /aa/ is in the carry-phrase /iy aa iy/, and vowel
/eh/ in /iy eh iy/. Note that as the speaking rate increases, the distinction between vowels /aa/ and /eh/
measured by the difference between their static formant values gradually diminishes. The same γ value
of 0.9 is used in generating all points in the figure

degree of “static” sound confusion as speaking rate increases is clearly evident from both the
measurement data (Fig. 5.6) and prediction (Fig. 5.5).

5.2.4 Model Prediction of Vocal Tract Resonance Trajectories for Real

Speech Utterances
We have used the expected VTR trajectories computed from the HTM to predict actual VTR
frequency trajectories for real speech utterances from the TIMIT database. Only the phone
identities and their boundaries are input to the model for the prediction, and no use is made of
speech acoustics. Given the phone sequence in any utterance, we first break up the compound
phones (affricates and diphthongs) into their constituents. Then we obtain the initial VTR
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 81

Data --- Speaker A (Pitermann, 2000)
2000

/ε/
1800

f2
1600
Ave. measured formant frequencies (Hz)

/a/
1400

1200

1000

800

/a/
600

f1
400

/ε/
200
40 50 60 70 80 90 100 110 120

Speaking rate (beat/min)

FIGURE 5.6: The formant measurement data from literature are reorganized and plotted, showing
similar trends to the model prediction under similar conditions

target values based on limited context dependency by table lookup (see details in [9], Ch. 13).
Then automatic and iterative target adaptation is performed for each phone-like unit based
on the difference between the results of a VTR tracker (described in [126]) and the VTR
prediction from the FIR filter model. These target values are provided not only to vowels, but
also to consonants for which the resonance frequency targets are used with weak or no acoustic
manifestation. The converged target values, together with the phone boundaries provided from
the TIMIT database, form the input to the FIR filter of the HTM and the output of the filter
gives the predicted VTR frequency trajectories.
Three example utterances from TIMIT (SI1039, SI1669 and SI2299) are shown in
Figs. 5.7–5.9. The stepwise dashed lines ( f 1 / f 2 / f 3 / f 4 ) are the target sequences as inputs to the
FIR filter, and the continuous lines ( f 1 / f 2 / f 3 / f 4 ) are the outputs of the filter as the predicted
VTR frequency trajectories. Parameters γ and D are fixed and not automatically learned. To
facilitate assessment of the accuracy in the prediction, the inputs and outputs are superimposed
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

82 DYNAMIC SPEECH MODELS

γ = [0.6], D = 7
6

4
Frequency (kHz)

0
0 50 100 150 200 250 300 350
Frame (10 ms)

FIGURE 5.7: The f 1 / f 2 / f 3 / f 4 VTR frequency trajectories (smooth lines) generated from the FIR model
for VTR target filtering using the phone sequence and duration of a speech utterance (SI1039) taken from
the TIMIT database. The target sequence is shown as stepwise lines, switching at the phone boundaries
labeled in the database. They are superimposed on the utterance’s spectrogram. The utterance is “He has
never, himself, done anything for which to be hated—which of us has ”

on the spectrograms of these utterances, where the true resonances are shown as the dark
bands. For the majority of frames, the filter’s output either coincides or is close to the true VTR
frequencies, even though no acoustic information is used. Also, comparing the input and output
of the filter, we observe only a rather mild degree of target undershooting or reduction in these
and many other TIMIT utterances we have examined but not shown here.

5.2.5 Simulation Results on Model Prediction for Cepstral Trajectories

The predicted VTR trajectories in Figs. 5.7–5.9 are fed into the nonlinear mapping function
in the HTM to produce the predicted linear cepstra shown in Figs. 5.10–5.12, respectively, for
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 83

γ = [0.6], D = 7
6

4
Frequency (kHz)

0
0 20 40 60 80 100 120 140 160 180 200
Frame (10 ms)

FIGURE 5.8: Same as Fig. 5.7 except with another utterance “Be excited and don’t identify yourself ”
(SI1669)

the three example TIMIT utterances. Note that the model prediction includes residual means,
which are trained from the full TIMIT data set using an HTK tool. The zero-mean random
component of the residual is ignored in these figures. The residual means for the substates
(three for each phone) are added sequentially to the output of the nonlinear function Eq. (5.12),
assuming each substate occupies three equal-length subsegments of the entire phone segment
length provided by TIMIT database. To avoid display cluttering, only linear cepstra with orders
one (C1), two (C2) and three (C3) are shown here, as the solid lines. Dashed lines are the linear
cepstral data C1, C2 and C3 computed directly from the waveforms of the same utterances
for comparison purposes. The data and the model prediction generally agree with each other,
somewhat better for lowerorder cepstra than for higherorder ones. It was found that these
discrepancies are generally within the variances of the prediction residuals automatically trained
from the entire TIMIT training set (using an HTK tool for monophone HMM training).
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

84 DYNAMIC SPEECH MODELS

γ = [0.6], D = 7
6

4
Frequency (kHz)

0
0 50 100 150 200 250
Frame (10 ms)

FIGURE 5.9: Same as Fig. 5.7 except with the third utterance “Sometimes, he coincided with my father’s
being at home ” (SI2299)

5.3 PARAMETER ESTIMATION

In this section, we will present in detail a novel parameter estimation algorithm we have devel-
oped and implemented for the HTM described in the preceding section, using the linear cepstra
as the acoustic observation data in the training set. The criterion used for this training is to
maximize the acoustic observation likelihood in Eq. (5.20). The full set of the HTM parameters
consists of those characterizing the linear cepstra residual distributions and those characterizing
the VTR target distributions. We present their estimation separately below, assuming that all
phone boundaries are given (as in the TIMIT training data set).

5.3.1 Cepstral Residuals’ Distributional Parameters

This subset of the HTM parameters consists of (1) the mean vectors μr s and (2) the diagonal
elements σr2s in the covariance matrices of the cepstral prediction residuals. Both of them are
conditioned on phone or sub-phone segmental unit s .
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 85

2

C1 0

−1

−2
0 50 100 150 200 250 300 350

C2 0

−1

0 50 100 150 200 250 300 350

0.5

C3
0

−0.5

0 50 100 150 200 250 300 350

Frame (10 ms)

FIGURE 5.10: Linear cepstra with order one (C1), two (C2) and three (C3) predicted from the final
stage of the model generating the linear cepstra (solid lines) with the input from the FIR filtered results
(for utterance SI1039). Dashed lines are the linear cepstral data C1, C2 and C3 computed directly from
the waveform

Mean Vectors
To find the ML (maximum likelihood) estimate of parameters μr s , we set

K
∂ log p(o(k) | s )
k=1
= 0,
∂μr s

where p(o(k) | s ) is given by Eq. (5.20), and K denotes the total duration of sub-phone s in the
training data. This gives


K 
o(k) − μ̄o s = 0, or (5.23)
k=1
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

86 DYNAMIC SPEECH MODELS

1
C1 0

−1

−2
0 20 40 60 80 100 120 140 160 180 200

0.5

C2 0

−0.5

−1
0 20 40 60 80 100 120 140 160 180 200

0.5

C3 0

−0.5

0 20 40 60 80 100 120 140 160 180 200

Frame (10 ms)

FIGURE 5.11: Same as Fig. 5.10 except with the second utterance (SI2299)


K
o(k) − F  [z 0 (k)]μz(k)
k=1 
−{F[z 0 (k)] + μr s − F  [z 0 (k)]z 0 (k)} = 0. (5.24)

Solving for μr s , we have the estimation formula of

 
k o(k) − F[z 0 (k)] − F  [z 0 (k)]μ z(k) + F  [z 0 (k)]z 0 (k)
μ̂r s = . (5.25)
K

Diagonal Covariance Matrices

Denote the diagonal elements of the covariance matrices for the residuals as a vector σr2s . To
derive the ML estimate, we set
K
∂ log k=1 p(o(k) | s )
= 0,
∂σr2s
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 87

2

1
C1
0

−1

−2
0 50 100 150 200 250

0.5
C2 0

−0.5

−1
0 50 100 150 200 250

0.5

C3 0

−0.5

0 50 100 150 200 250

Frame (10 ms)

FIGURE 5.12: Same as Fig. 5.10 except with the third utterance (SI1669)

which gives


K
σr2s + q(k) − (o(k) − μ̄o s )2 
= 0, (5.26)
k=1
[σr2s + q(k)]2

where vector squaring is the element-wise operation, and


q(k) = diag F  [z0 (k)]Σz(k)(F  [z0 (k)])Tr . (5.27)

Due to the frame (k) dependency in the denominator in Eq. (5.26), no simple closed-form
solution is available for solving σr2s from Eq. (5.26). We have implemented three different
techniques for seeking approximate ML estimates that are outlined as follows:

1. Frame-independent approximation: Assume the dependency of q(k) on time frame k is

mild, or q(k) ≈ q̄. Then the denominator in Eq. (5.26) can be cancelled, yielding the
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

88 DYNAMIC SPEECH MODELS

approximate closed-form estimate of
K  
k=1 (o(k) − μ̄o s )2 − q(k)
σ̂r2s ≈ . (5.28)
K
2. Direct gradient ascent: Make no assumption of the above, and take the left-hand side of
Eq. (5.26) as the gradient ∇ L of log-likelihood of the data in the standard gradient-
ascent algorithm:

σr2s (t + 1) = σr2s (t) + t ∇L(o 1K | σr2s (t)),

where t is a heuristically chosen positive constant controlling the learning rate at the
t-th iteration.
3. Constrained gradient ascent: Add to the previous standard gradient ascent technique the
constraint that the variance estimate be always positive. The constraint is established
by the parameter transformation: σ̃r2s = log σr2s , and by performing gradient ascent for
σ̃r2s instead for σr2s :

σ̃r2s (t + 1) = σ̃r2s (t) + ˜t ∇ L̃(o 1K | σ̃r2s (t)),

Using chain rule, we show below that the new gradient ∇ L̃ is related to the gradient
∇ L before parameter transformation in a simple manner:

∂L̃ ∂ L̃ ∂σr2s
∇ L̃ = = = (∇L) exp(σ̃r2s ).
∂ σ̃r2s ∂σr2s ∂ σ̃r2s

At the end of the algorithm iteration, the parameters are transformed via σr2s = exp(σ̃r2s ),
which is guaranteed to be positive.

For efficiency purposes, parameter updating in the above gradient ascent techniques is
carried out after each utterance in the training, rather than after the entire batch of all utterances.
We note that the quality of the estimates for the residual parameters discussed above plays
a crucial role in phonetic recognition performance. These parameters provide an important
mechanism for distinguishing speech sounds that belong to different manners of articulation.
This is attributed to the fact that nonlinear cepstral prediction from VTRs has different accuracy
for these different classes of sounds. Within the same manner class, the phonetic separation
is largely accomplished by distinct VTR targets, which typically induce significantly different
cepstral prediction values via the “amplification” mechanism provided by the Jacobian matrix
F  [z].
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 89

5.3.2 Vocal Tract Resonance Targets’ Distributional Parameters
This subset of the HTM parameters consists of (1) the mean vectors μTs and (2) the diagonal
elements σ 2Ts in the covariance matrices of the stochastic segmental VTR targets. They are also
conditioned on phone segment s (and not on sub-phone segment).

Mean Vectors
To obtain a closed-form estimation solution, we assume diagonality of the prediction cepstral
residual’s covariance matrix Σr s . Denoting its q th component by σr2 (q ) (q = 1, 2, . . . , Q), we
decompose the multivariate Gaussian of Eq. (5.20) element-by-element into


J
1  (o k ( j ) − μ̄o ( j ))2 
p(o(k) | s (k)) =  exp − 2
s (k)
, (5.29)
j =1 2πσo2s (k) ( j ) 2σ o s (k) j )
(

where o k ( j ) denotes the j th component (i.e., j th order) of the cepstral observation vector at
frame k.
The log-likelihood function for a training data sequence (k = 1, 2, . . . , K ) relevant to
the VTR mean vector μTs becomes

J 
K 
(o k ( j ) − μ̄o s (k) ( j ))2 
P= − (5.30)
k=1 j =1
σo2s (k) ( j )
J [
  2
K  
f F [z0 (k), j, f ] l a k (l)μT (l, f ) − d k ( j )]
= ,
k=1 j =1
σo2s (k) ( j )

where l and f are indices to phone and to VTR component, respectively, and

dk ( j ) = o k ( j ) − F[z0 (k), j ] + F  [z0 (k), j, f ]z0 (k, f ) − μr s (k) ( j ).
f

While the acoustic feature’s distribution is Gaussian for both HTM and HMM given
the state s , the key difference is that the mean and variance in HTM as in Eq. (5.20) are both
time-varying functions (hence trajectory model). These functions provide context dependency
(and possible target undershooting) via the smoothing of targets across phonetic units in the
utterance. This smoothing is explicitly represented in the weighted sum over all phones in the

utterance (i.e., l ) in Eq. (5.30).
Setting

∂P
= 0,
∂μT (l 0 , f 0 )
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

90 DYNAMIC SPEECH MODELS

and grouping terms involving unknown μT (l, f ) on the left and the remaining terms on the
right, we obtain

A(l, f ; l 0 , f 0 )μT (l, f )
f l
 F  [z0 (k), j, f 0 ] 
= d k ( j ) a k (l 0 ), (5.31)
k j
σo2s (k) ( j )

with f 0 = 1, 2, . . . , 8 for each VTR dimension, and with l 0 = 1, 2, . . . , 58 for each phone unit.
In Eq. (5.31),

 F  [z0 (k), j, f ]F  [z0 (k), j, f 0 ]

A(l, f ; l 0 , f 0 ) = a k (l 0 )a k (l). (5.32)
k, j
σ 2 ( j)
o s (k)

Eq. (5.31) is a 464 × 464 full-rank linear system of equations. (The dimension 464 =
58 × 8 where we have a total of 58 phones in the TIMIT database after decomposing each
diphthong into two “phones”, and 8 is the VTR vector dimension.) Matrix inversion gives an
ML estimate of the complete set of target mean parameters: a 464-dimensional vector formed
by concatenating all eight VTR components (four frequencies and four bandwidths) of the
58 phone units in TIMIT.
In implementing Eq. (5.31) for the ML solution to target mean vectors, we kept other
model parameters constant. Estimation of the target and residual parameters was carried out
in an iterative manner. Initialization of the parameters μT (l, f ) was provided by the values
described in [9].
An alternative training of the target mean parameters in a simplified version of the HTM
and its experimental evaluation are described in [112]. In that training, the VTR tracking
results obtained by the tracking algorithm described in Chapter 4 are exploited as the basis for
learning, contrasting the learning described in this section, which uses the raw cepstral acoustic
data only. Use of the VTR tracking results enables speaker-adaptive learning for the VTR target
parameters as shown in [112].

Diagonal Covariance Matrices

To establish the objective function for optimization, we take logarithm on the sum of the
likelihood function Eq. (5.29) (over K frames) to obtain
 
K  J
(o k ( j ) − μ̄o s (k) ( j ))2
LT ∝ − + log [σr2s ( j ) + q (k, j )] , (5.33)
k=1 j =1
σ 2 ( j ) + q (k, j )
rs
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 91

where q (k, j ) is the j th element of the vector q(k) as defined in Eq. (5.27). When Σz(k) is
diagonal, it can be shown that
 
q (k, j ) = σz(k)
2
( f )(F j f )2 = vk (l)σT2 (l, f )(F j f )2 , (5.34)
f f l

where F j f is the ( j, f ) element of Jacobian matrix F  [·] in Eq. (5.27), and the second equality
is due to Eq. (5.11).
Using chain rule to compute the gradient, we obtain

∂ LT
∇ LT (l, f ) = (5.35)
∂σT2 (l, f )
 
K  J (o k ( j ) − μ̄o s (k) ( j ))2 (F j f )2 vk (l) (F j f )2 vk (l)
= − .
k=1 j =1
[σr2s ( j ) + q (k, j )]2 σr2s ( j ) + q (k, j )

Gradient-ascend iterations then proceed as follows:

σ 2T (l, f ) ← σ 2T (l, f ) + ∇ LT (l, f ),

for each phone l and for each element f in the diagonal VTR target covariance matrix.

5.4 APPLICATION TO PHONETIC RECOGNITION

5.4.1 Experimental Design
Phonetic recognition experiments have been conducted [124] aimed at evaluating the HTM
and the parameter learning algorithms described in this chapter. The standard TIMIT phone
set with 48 labels is expanded to 58 (as described in [9]) in training the HTM parameters using
the standard training utterances. Phonetic recognition errors are tabulated using the commonly
adopted 39 labels after the label folding. The results are reported on the standard core test set
of 192 utterances by 24 speakers [127].
Due to the high implementation and computational complexity for the full-fledged HTM
decoder, the experiments reported in [124] have been restricted to those obtained by N-best
rescoring and lattice-constrained search. For each of the core test utterances, a standard decision-
tree-based triphone HMM with the bi-gram language model is used to generate a large N-best
list (N = 1000) and a large lattice. These N-best lists and lattices are used for the rescoring
experiments with the HTM. The HTM system is trained using the parameter estimation
algorithms described earlier in this chapter. Learning rates in the gradient ascent techniques
have been tuned empirically.
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

92 DYNAMIC SPEECH MODELS

5.4.2 Experimental Results
In Table 5.1, phonetic recognition performance comparisons are shown between the HMM
system described above and three evaluation versions of the HTM system. The HTM-1 version
uses the HTM likelihood computed from Eq. (5.20) to rescore the 1000-best lists, and no HMM
score and language model (LM) score attached in the 1000-best list are exploited. The HTM-2
version improves the HTM-1 version slightly by linearly weighting the log-likelihoods of the
HTM, the HMM and the (bigram) LM, based on the same 1000-best lists. The HTM-3
version replaces the 1000-best lists by the lattices, and carries out A* search, constrained by the
lattices and with linearly weighted HTM–HMM–LM scores, to decode phonetic sequences.
(See a detailed technical description of this A*-based search algorithm in [111].) Notable
performance improvement is obtained as shown in the final row of Table 5.1. For all the
systems, the performance is measured by percent phone recognition accuracy (i.e., including
insertion errors) averaged over the core test-set sentences (numbers in bolds in column 2). The
percent-correctness performance (i.e., excluding insertion errors) is listed in column 3. The
substitution, deletion and insertion error rates are shown in the remaining columns.
The performance results in Table 5.1 are obtained using the identical acoustic features of
frequency-warped linear cepstra for all the systems. Frequency warping of linear cepstra [128]
has been implemented by a linear matrix-multiplication technique on both acoustic features and
the observation-prediction component of the HTM. The warping gives slight performance
improvement for both HMM and HTM systems by a similar amount. Overall, the lattice-
based HTM system (75.07% accuracy) gives 13% fewer errors than does the HMM system

TABLE5.1: TIMIT Phonetic Recognition Performance Comparisons Between

an HMM System and Three Versions of the HTM System

ACC % CORR % SUB % DEL % INS %

HMM 71.43 73.64 17.14 9.22 2.21
HTM-1 74.31 77.76 16.23 6.01 3.45
HTM-2 74.59 77.73 15.61 6.65 3.14
HTM-3 75.07 78.28 15.94 5.78 3.20
Note. HTM-1: N-best rescoring with HTM scores only; HTM-2: N-
best rescoring with weighted HTM, HMM and LM scores; HTM-3:
Lattice-constrained A* search with weighted HTM, HMM and LM
scores. Identical acoustic features (frequency-warped linear cepstra) are
used.
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

MODELS WITH CONTINUOUS-VALUED HIDDEN SPEECH TRAJECTORIES 93

(71.43% accuracy). This performance is better than that of any HMM system on the same task
as summarized in [127], and is approaching the best-ever result (75.6% accuracy) obtained by
using many heterogeneous classifiers as reported in [127] also.

5.5 SUMMARY
In this chapter, we present in detail a second specific type of hidden dynamic models, which
we call the hidden trajectory model (HTM). The unique character of the HTM is that the
hidden dynamics are represented not by temporal recursion on themselves but by explicit “tra-
jectories” or hidden trended functions constructed by FIR filtering of targets. In contrast to
the implementation strategy for the model discussed in Chapter 4 where the hidden dynamics
are discretized, the implementation strategy in the HTM maintains continuous-valued hidden
dynamics, and introduces approximations by constraining the temporal boundaries associated
with discrete phonological states. Given such constraints, rigorous algorithms for model param-
eter estimation are developed and presented without the need to approximate the continuous
hidden dynamic variables by their discretized values as done in Chapter 4.
The main portions of this chapter are devoted to formal construction of the HTM, its
computer simulation and the parameter estimation algorithm’s development. The computa-
tionally efficient decoding algorithms have not been presented, as they are still under research
and development and are hence not appropriate to describe in this book at present. In contrast,
decoding algorithms for discretized hidden dynamic models are much more straightforward to
develop, as we have presented in Chapter 4.
Although we present only two types of implementation strategies in this book (Chapters
4, 5, respectively) for dynamic speech modeling within the general computational framework
established in Chapter 2, other types of implementation strategies and approximations (such as
variational learning and decoding) are possible. We have given some related references at the
beginning of this chapter.
As a summary and conclusion of this book, we have provided scientific background, math-
ematical theory, computational framework, algorithmic development and technological needs
and two selected applications for dynamic speech modeling, which is the theme of this book.
A comprehensive survey in this area of research is presented, drawing on the work of a number
of (non-exhaustive) research groups and individual researchers worldwide. This direction of
research is guided by scientific principles applied to study human speech communication, and is
based on the desire to acquire knowledge about the realistic dynamic process in the closed-loop
speech chain. It is hoped that with integration of this unique style of research with other pow-
erful pattern recognition and machine learning approaches, the dynamic speech models, as they
become better developed, will form a foundation for the next-generation speech technology
serving the humankind and society.
P1: IML/FFX P2: IML
MOBK024-05 MOBK024-LiDeng.cls April 26, 2006 14:3

94
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

95

Bibliography
[1] P. Denes and E. Pinson. The Speech Chain, 2nd edn, Worth Publishers, New York, 1993.
[2] K. Stevens. Acoustic Phonetics, MIT Press, Cambridge, MA, 1998.
[3] K. Stevens. “Toward a model for lexical access based on acoustic landmarks and distinc-
tive features,” J. Acoust. Soc. Am., Vol. 111, April 2002, pp. 1872–1891.
[4] L. Rabiner and B.-H. Juang. Fundamentals of Speech Recognition, Prentice-Hall, Upper
Saddle River, NJ, 1993.
[5] X. Huang, A. Acero, and H. Hon. Spoken Language Processing, Prentice Hall, New York,
2001.
[6] V. Zue. “Notes on speech spectrogram reading,” MIT Lecture Notes, Cambridge, MA,
1991.
[7] J. Olive, A. Greenwood, and J. Coleman. Acoustics of American English Speech—A Dy-
namic Approach, Springer-Verlag, New York, 1993.
[8] C. Williams. “How to pretend that correlated variables are independent by using dif-
ference observations,” Neural Comput., Vol. 17, 2005, pp. 1–6.
[9] L. Deng and D. O’Shaughnessy. SPEECH PROCESSING—A Dynamic and
Optimization-Oriented Approach (ISBN: 0-8247-4040-8), Marcel Dekker, New York,
2003, pp. 626.
[10] L. Deng and X.D. Huang. “Challenges in adopting speech recognition,” Commun.
ACM, Vol. 47, No. 1, January 2004, pp. 69–75.
[11] M. Ostendorf. “Moving beyond the beads-on-a-string model of speech,” in Proceedings
of IEEE Workshop on Automatic Speech Recognition and Understanding, December 1999,
Keystone, co, pp. 79–83.
[12] N. Morgan, Q. Zhu, A. Stolcke, et al. “Pushing the envelope—Aside,” IEEE Signal
Process. Mag., Vol. 22, No. 5, September. 2005, pp. 81–88.
[13] F. Pereira. “Linear models for structure prediction,” in Proceedings of Interspeech, Lisbon,
September 2005, pp. 717–720.
[14] M. Ostendorf, V. Digalakis, and J. Rohlicek. “From HMMs to segment models: A
unified view of stochastic modeling for speech recognition” IEEE Trans. Speech Audio
Process., Vol. 4, 1996, pp. 360–378.
[15] B.-H. Juang and S. Katagiri. “Discriminative learning for minimum error classification,”
IEEE Trans. Signal Process., Vol. 40, No. 12, 1992, pp. 3043–3054.
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

96 DYNAMIC SPEECH MODELS

[16] D. Povey. “Discriminative training for large vocabulary speech recognition,” Ph.D. dis-
sertation, Cambridge University, 2003.
[17] W. Chou and B.-H. Juang (eds.). Pattern Recognition in Speech and Language Processing,
CRC Press, Boca Raton, FL, 2003.
[18] L. Deng, J. Wu, J. Droppo, and A. Acero. “Analysis and comparison of two feature
extraction/compensation algorithms,” IEEE Signal Process. Lett., Vol. 12, No. 6, June
2005, pp. 477–480.
[19] D. Povey, B. Kingsbury, L. Mangu, G. Saon, H. Solatu, and G. Zweig. “FMPE: Dis-
criminatively trained features for speech recognition,” IEEE Proc. ICASSP, Vol. 2, 2005,
pp. 961–964.
[20] J. Bilmes and C. Bartels. “Graphical model architectures for speech recognition,” IEEE
Signal Process. Mag., Vol. 22, No. 5, Sept. 2005, pp. 89–100.
[21] G. Zweig. “Bayesian network structures and inference techniques for automatic speech
recognition,” Comput. Speech Language, Vol. 17, No. 2/3, 2003, pp. 173–193.
[22] F. Jelinek, et al. “Central issues in the recognition of conversational speech,” Summary
Report, Johns Hopkins University, Baltimore, MD, 1999, pp. 1–57.
[23] S. Greenberg, J. Hollenback, and D. Ellis. “Insights into spoken language gleaned from
phonetic transcription of the Switchboard corpus,” Proc. ICSLP, Vol. 1, 1996, pp. S32–
S35.
[24] L. Deng and J. Ma. “Spontaneous speech recognition using a statistical coarticulatory
model for the hidden vocal–tract–resonance dynamics,” J. Acoust. Soc. Am., Vol. 108,
No. 6, 2000, pp. 3036–3048.
[25] S. Furui, K. Iwano, C. Hori, T. Shinozaki, Y. Saito, and S. Tamur. “Ubiquitous speech
processing,” IEEE Proc. ICASSP, Vol. 1, 2001, pp. 13–16.
[26] K.C. Sim and M. Gales. “Temporally varying model parameters for large vocabulary
continuous speech recognition,” in Proceedings of Interspeech, Lisbon, September 2005,
pp. 2137–2140.
[27] K.-F. Lee. Automatic speech recognition: The Development of the Sphinx Recognition System,
Springer, New York, 1988.
[28] C.-H. Lee, F. Soong, and K. Paliwal (eds.). Automatic Speech and Speaker Recognition—
Advanced Topics, Kluwer Academic, Norwell, MA, 1996.
[29] F. Jelinek. Statistical Methods for Speech Recognition, MIT Press, Cambridge, MA, 1997.
[30] B.-H. Juang and S. Furui (Eds.). Proc. IEEE (special issue), Vol. 88, 2000.
[31] L. Deng, K. Wang, and W. Chou. “Speech technology and systems in human–Machine
communication—Guest editors’ editorial,” IEEE Signal Process. Mag., Vol. 22, No. 5,
September 2005, pp. 12–14.
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

BIBLIOGRAPHY 97
[32] J. Allen. “How do humans process and recognize speech,” IEEE Trans. Speech Audio
Process., Vol. 2, 1994, pp. 567–577.
[33] L. Deng. “A dynamic, feature-based approach to the interface between phonology and
phonetics for speech modeling and recognition,” Speech Commun., Vol. 24, No. 4, 1998,
pp. 299–323.
[34] H. Bourlard, H. Hermansky, and N. Morgan. “Towards increasing speech recognition
error rates,” Speech Commun., Vol. 18, 1996, pp. 205–231.
[35] L. Deng. “Switching dynamic system models for speech articulation and acoustics,”
in M. Johnson, M. Ostendorf, S. Khudanpur, and R. Rosenfeld (eds.), Mathemati-
cal Foundations of Speech and Language Processing, Springer-Verlag, New York, 2004,
pp. 115–134.
[36] R. Lippmann. “Speech recognition by human and machines,” Speech Commun., Vol. 22,
1997, pp. 1–14.
[37] L. Pols. “Flexible human speech recognition,” in Proceedings of the IEEE Workshop on
Automatic Speech Recognition and Understanding, 1997, Santa Barbara, CA, pp. 273–283.
[38] C.-H. Lee. “From knowledge-ignorant to knowledge-rich modeling: A new speech
research paradigm for next-generation automatic speech recognition,” in Proc. ICSLP,
Jeju Island, Korea, October 2004, pp. 109–111.
[39] M. Russell. “Progress towards speech models that model speech,” in Proc. IEEE Workshop
on Automatic Speech Recognition and Understanding, 1997, Santa Barbara, CA, pp. 115–
123.
[40] M. Russell. “A segmental HMM for speech pattern matching,” IEEE Proceedings of the
ICASSP, Vol. 1, 1993, pp. 499–502.
[41] L. Deng. “A generalized hidden Markov model with state-conditioned trend functions
of time for the speech signal,” Signal Process., Vol. 27, 1992, pp. 65–78.
[42] J. Bridle, L. Deng, J. Picone, et al. “An investigation of segmental hidden dynamic
models of speech coarticulation for automatic speech recognition,” Final Report for the
1998 Workshop on Language Engineering, Center for Language and Speech Processing
at Johns Hopkins University, 1998, pp. 1–61.
[43] K. Kirchhoff. “Robust speech recognition using articulatory information,” Ph.D. thesis,
University of Bielfeld, Germany, July 1999.
[44] R. Bakis. “Coarticulation modeling with continuous-state HMMs,” in Proceedings
of the IEEE Workshop on Automatic Speech Recognition, Harriman, New York, 1991,
pp. 20–21.
[45] Y. Gao, R. Bakis, J. Huang, and B. Zhang. “Multistage coarticulation model combining
articulatory, formant and cepstral features,” Proc. ICSLP, Vol. 1, 2000, pp. 25–28.
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

98 DYNAMIC SPEECH MODELS

[46] J. Frankel and S. King. “ASR—Articulatory speech recognition,” Proc. Eurospeech, Vol. 1,
2001, pp 599–602.
[47] T. Kaburagi and M. Honda. “Dynamic articulatory model based on multidimensional
invariant-feature task representation,” J. Acoust. Soc. Am., 2001, Vol. 110, No. 1, pp. 441–
452.
[48] P. Jackson, B. Lo, and M. Russell. “Data-driven, non-linear, formant-to-acoustic map-
ping for ASR,” IEE Electron. Lett., Vol. 38, No. 13, 2002, pp. 667–669.
[49] M. Russell and P. Jackson. “A multiple-level linear/linear segmental HMM with a
formant-based intermediate layer,” Comput. Speech Language, Vol. 19, No. 2, 2005,
pp. 205–225.
[50] L. Deng and D. Sun. “A statistical approach to automatic speech recognition using the
atomic speech units constructed from overlapping articulatory features,” J. Acoust. Soc.
Am., Vol. 95, 1994, pp. 2702–2719.
[51] H. Nock and S. Young. “Loosely coupled HMMs for ASR: A preliminary study,”
Technical Report TR386, Cambridge University, 2000.
[52] K. Livescue, J. Glass, and J. Bilmes. “Hidden feature models for speech recognition
using dynamic Bayesian networks,” Proc. Eurospeech, Vol. 4, 2003, pp. 2529–2532.
[53] E. Saltzman and K. Munhall. “A dynamical approach to gestural patterning in speech
production,” Ecol. Psychol., Vol. 1, pp. 333–382.
[54] L. Deng. “Computational models for speech production,” in K. Ponting (ed.), Com-
putational Models of Speech Pattern Processing (NATO ASI Series), Springer, New York,
1999, pp. 199–214.
[55] L. Deng, M. Aksmanovic, D. Sun, and J. Wu. “Speech recognition using hidden Markov
models with polynomial regression functions as nonstationary states,” IEEE Trans.
Speech Audio Process., Vol. 2, 1994, pp. 507–520.
[56] C. Li and M. Siu, “An efficient incremental likelihood evaluation for polynomial
trajectory model with application to model training and recognition,” IEEE Proc.
ICASSP, Vol. 1, 2003, pp. 756–759.
[57] Y. Minami, E. McDermott, A. Nakamura, and S. Katagiri. “Recognition method with
parametric trajectory generated from mixture distribution HMMs,” IEEE Proc. ICASSP,
Vol. 1, 2003, pp. 124–127.
[58] C. Blackburn and S. Young. “A self-learning predictive model of articulator move-
ments during speech production,” J. Acoust. Soc. Am., Vol. 107, No. 3, 2000, pp. 1659–
1670.
[59] L. Deng, G. Ramsay, and D. Sun. “Production models as a structural basis for automatic
speech recognition,” Speech Commun., Vol. 22, No. 2, 1997, pp. 93–111.
[60] B. Lindblom. “Explaining phonetic variation: A sketch of the H & H theory,” in
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

BIBLIOGRAPHY 99
W. Hardcastle and A. Marchal (eds.), Speech Production and Speech Modeling, Kluwer,
Norwell, MA, 1990, pp. 403–439.
[61] N. Chomsky and M. Halle. The Sound Pattern of English, Harper and Row, New York,
1968.
[62] N. Clements. “The geometry of phonological features,” Phonology Yearbook, Vol. 2, 1985,
pp. 225–252.
[63] C. Browman and L. Goldstein. “Articulatory phonology: An overview,” Phonetica,
Vol. 49, 1992, pp. 155–180.
[64] M. Randolph. “Speech analysis based on articulatory behavior,” J. Acoust. Soc. Am.,
Vol. 95, 1994, p. 195.
[65] L. Deng and H. Sameti. “Transitional speech units and their representation by the
regressive Markov states: Applications to speech recognition,” IEEE Trans. Speech Audio
Process., Vol. 4, No. 4, July 1996, pp. 301–306.
[66] J. Sun, L. Deng, and X. Jing. “Data-driven model construction for continuous speech
recognition using overlapping articulatory features,” Proc. ICSLP, Vol. 1, 2000, pp. 437–
440.
[67] Z. Ghahramani and M. Jordan. “Factorial hidden Markov models,” Machine Learn.,
Vol. 29, 1997, pp.245–273.
[68] K. Stevens. “On the quantal nature of speech,” J. Phonetics, Vol. 17, 1989, pp. 3–45.
[69] A. Liberman and I. Mattingly. “The motor theory of speech perception revised,” Cog-
nition, Vol. 21, 1985, pp. 1–36.
[70] B. Lindblom. “Role of articulation in speech perception: Clues from production,”
J. Acoust. Soc. Am., Vol. 99, No. 3, 1996, pp. 1683–1692.
[71] P. MacNeilage. “Motor control of serial ordering in speech,” Psychol. Rev., Vol. 77, 1970,
pp. 182–196.
[72] R. Kent, G. Adams, and G. Turner. “Models of speech production,” in N. Lass (ed.),
Principles of Experimental Phonetics, Mosby, London, 1995, pp. 3–45.
[73] J. Perkell, M. Matthies, M. Svirsky, and M. Jordan. “Goal-based speech motor con-
trol: A theoretical framework and some preliminary data,” J. Phonetics, Vol. 23, 1995,
pp. 23–35.
[74] J. Perkell. “Properties of the tongue help to define vowel categories: Hypotheses based
on physiologically-oriented modeling,” J. Phonetics, Vol. 24, 1996, pp. 3–22.
[75] P. Perrier, D. Ostry, and R. Laboissière. “The equilibrium point hypothesis and its
application to speech motor control,” J. Speech Hearing Res., Vol. 39, 1996, pp. 365–378.
[76] B. Lindblom, J. Lubker, and T. Gay. “Formant frequencies of some fixed-mandible
vowels and a model of speech motor programming by predictive simulation,” J. Phonetics,
Vol. 7, 1979, pp. 146–161.
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

100 DYNAMIC SPEECH MODELS

[77] S. Maeda. “On articulatory and acoustic variabilities,” J. Phonetics, Vol. 19, 1991, pp. 321–
331.
[78] G. Ramsay and L. Deng. “A stochastic framework for articulatory speech recognition,”
J. Acoust. Soc. Am., Vol. 95, No. 6, 1994, p. 2871.
[79] C. Coker. “A model of articulatory dynamics and control,” Proc. IEEE, Vol. 64, No. 4,
1976, pp. 452–460.
[80] P. Mermelstein. “Articulatory model for the study of speech production,” J. Acoust. Soc.
Am., Vol. 53, 1973, pp. 1070–1082.
[81] C. Bishop. Neural Networks for Pattern Recognition, Clarendon Press, Oxford, 1995.
[82] Z. Ghahramani and S. Roweis. “Learning nonlinear dynamic systems using an EM
algorithm,” Adv. Neural Informat. Process. Syst., Vol. 11, 1999, pp. 1–7.
[83] L. Deng, J. Droppo, and A. Acero. “Estimating cepstrum of speech under the presence
of noise using a joint prior of static and dynamic features,” IEEE Trans. Speech Audio
Process., Vol. 12, No. 3, May 2004, pp. 218–233.
[84] J. Ma and L. Deng. “Target-directed mixture linear dynamic models for spontaneous
speech recognition,” IEEE Trans. Speech Audio Process., Vol. 12, No. 1, 2004, pp. 47–58.
[85] J. Ma and L. Deng. “A mixed-level switching dynamic system for continuous speech
recognition,” Comput. Speech Language, Vol. 18, 2004, pp. 49–65.
[86] H. Gish and K. Ng. “A segmental speech model with applications to word spotting,”
IEEE Proc. ICASSP, Vol. 1, 1993, pp. 447–450.
[87] L. Deng and M. Aksmanovic. “Speaker-independent phonetic classification using hid-
den Markov models with mixtures of trend functions,” IEEE Trans. Speech Audio Process.,
Vol. 5, 1997, pp. 319–324.
[88] H. Hon and K. Wang. “Unified frame and segment based models for automatic speech
recognition,” IEEE Proc. the ICASSP, Vol. 2, 2000, pp. 1017–1020.
[89] M. Gales and S. Young. “Segmental HMMs for speech recognition,” Proc. Eurospeech,
Vol. 3, 1993, pp. 1579–1582.
[90] W. Holmes and M. Russell. “Probabilistic-trajectory segmental HMMs,” Comput.
Speech Language, Vol. 13, 1999, pp. 3–27.
[91] C. Rathinavelu and L. Deng. “A maximum a posteriori approach to speaker adaptation
using the trended hidden Markov model,” IEEE Trans. Speech Audio Process., Vol. 9,
2001, pp. 549–557.
[92] O. Ghitza and M. Sondhi. “Hidden Markov models with templates as nonstationary
states: An application to speech recognition,” Comput. Speech Language, Vol. 7, 1993,
pp. 101–119.
[93] P. Kenny, M. Lennig, and P. Mermelstein. “A linear predictive HMM for vector-valued
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

BIBLIOGRAPHY 101
observations with applications to speech recognition,” IEEE Trans. Acoust., Speech, Signal
Process., Vol. 38, 1990, pp. 220–225.
[94] L. Deng and C. Rathinavalu. “A Markov model containing state-conditioned second-
order nonstationarity: Application to speech recognition,” Comput. Speech Language,
Vol. 9, 1995, pp. 63–86.
[95] A. Poritz. “Hidden Markov models: A guided tour,” IEEE Proc. ICASSP, Vol. 1, 1988,
pp. 7–13.
[96] H. Sheikhazed and L. Deng. “Waveform-based speech recognition using hidden filter
models: Parameter selection and sensitivity to power normalization,” IEEE Trans. Speech
Audio Process., Vol. 2, 1994, pp. 80–91.
[97] H. Zen, K. Tokuda, and T. Kitamura. “A Viterbi algorithm for a trajectory model derived
from HMM with explicit relationship between static and dynamic features,” IEEE Proc.
ICASSP, 2004, pp. 837–840.
[98] K. Tokuda, H. Zen, and T. Kitamura. “Trajectory modeling based on HMMs with the
explicit relationship between static and dynamic features,” Proc. Eurospeech, Vol. 2, 2003,
pp. 865–868.
[99] J. Tebelskis and A. Waibel. “Large vocabulary recognition using linked predictive neural
networks,” IEEE Proc. ICASSP, Vol. 1, 1990, pp. 437–440.
[100] E. Levin. “Word recognition using hidden control neural architecture,” IEEE Proc.
ICASSP, Vol. 1, 1990, pp. 433–436.
[101] L. Deng, K. Hassanein, and M. Elmasry. “Analysis of correlation structure for a neural
predictive model with application to speech recognition,” Neural Networks, Vol. 7, No. 2,
1994, pp. 331–339.
[102] V. Digalakis, J. Rohlicek, and M. Ostendorf. “ML estimation of a stochastic linear
system with the EM algorithm and its application to speech recognition,” IEEE Trans.
Speech Audio Process., Vol. 1, 1993, pp. 431–442.
[103] L. Deng. “Articulatory features and associated production models in statistical speech
recognition,” in K. Ponting (ed.), Computational Models of Speech Pattern Processing
(NATO ASI Series), Springer, New York, 1999, pp. 214–224.
[104] L. Lee, P. Fieguth, and L. Deng. “A functional articulatory dynamic model for speech
production,” IEEE Proc. ICASSP, Vol. 2, 2001, pp. 797–800.
[105] R. McGowan. “Recovering articulatory movement from formant frequency trajectories
using task dynamics and a genetic algorithm: Preliminary model tests,” Speech Commun.,
Vol. 14, 1994, pp. 19–48.
[106] R. McGowan and A. Faber. “Speech production parameters for automatic speech recog-
nition,” J. Acoust. Soc. Am., Vol. 101, 1997, p. 28.
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

102 DYNAMIC SPEECH MODELS

[107] J. Picone, S. Pike, R. Reagan, T. Kamm, J. Bridle, L. Deng, Z. Ma, H. Richards, and
M. Schuster. “Initial evaluation of hidden dynamic models on conversational speech,”
IEEE Proc. ICASSP, Vol. 1, 1999, pp 109–112.
[108] R. Togneri and L. Deng. “Joint state and parameter estimation for a target-directed non-
linear dynamic system model,” IEEE Trans. Signal Process., Vol. 51, No. 12, December
2003, pp. 3061–3070.
[109] L. Deng, D. Yu, and A. Acero. “A bi-directional target-filtering model of speech coar-
ticulation and reduction: Two-stage implementation for phonetic recognition,” IEEE
Trans. Speech Audio Process., Vol. 14, No. 1, Jan. 2006, pp. 256–265.
[110] L. Deng, A. Acero, and I. Bazzi. “Tracking vocal tract resonances using a quantized
nonlinear function embedded in a temporal constraint,” IEEE Trans. Speech Audio Pro-
cess., Vol. 14, No. 2, March 2006, pp. 425–434.
[111] D. Yu, L. Deng, and A. Acero. “Evaluation of a long-contextual-span trajectory model
and phonetic recognizer using A∗ lattice search,” in Proceedings of Interspeech, Lisbon,
September 2005, Vol. 1, pp. 553–556.
[112] D. Yu, L. Deng, and A. Acero. “Speaker-adaptive learning of resonance targets in a
hidden trajectory model of speech coarticulation,” Comput. Speech Language, 2006.
[113] H.B. Richards, and J.S. Bridle. “The HDM: A segmental hidden dynamic model of
coarticulation,” IEEE Proc. ICASSP, Vol. 1, 1999, pp. 357–360.
[114] F. Seide, J. Zhou, and L. Deng. “Coarticulation modeling by embedding a target-
directed hidden trajectory model into HMM—MAP decoding and evaluation,” IEEE
Proc. ICASSP, Vol. 2, 2003, pp. 748–751.
[115] L. Deng, X. Li, D. Yu, and A. Acero. “A hidden trajectory model with bi-directional
target-filtering: Cascaded vs. integrated implementation for phonetic recognition,”
IEEE Proceedings of the ICASSP, Philadelphia, 2005, pp. 337–340.
[116] L. Deng, D. Yu, and A. Acero. “Learning statistically characterized resonance targets
in a hidden trajectory model of speech coarticulation and reduction,” Proceedings of the
Eurospeech, Lisbon, 2005, pp. 1097–1100.
[117] L. Deng, I. Bazzi, and A. Acero. “Tracking vocal tract resonances using an analytical
nonlinear predictor and a target-guided temporal constraint,” Proceedings of the Eu-
rospeech, Vol. I, Geneva, Switzerland, September 2003, pp. 73–76.
[118] R. Togneri and L. Deng. “A state–space model with neural-network prediction for re-
covering vocal tract resonances in fluent speech from Mel-cepstral coefficients,” Comput.
Speech Language, 2006.
[119] A. Acero. “Formant analysis and synthesis using hidden Markov models,” in Proceedings
of the Eurospeech, Budapest, September 1999.
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

BIBLIOGRAPHY 103
[120] C. Huang and H. Wang. “Bandwidth-adjusted LPC analysis for robust speech recog-
nition,” Pattern Recognit. Lett., Vol. 24, 2003, pp. 1583–1587.
[121] L. Lee, H. Attias, and L. Deng. “Variational inference and learning for segmental
switching state space models of hidden speech dynamics,” in IEEE Proceedings of the
ICASSP, Vol. I, Hong Kong, April 2003, pp. 920–923.
[122] L. Lee, L. Deng, and H. Attias. “A multimodal variational approach to learning and
inference in switching state space models,” in IEEE Proceedings of the ICASSP, Montreal,
Canada, May 2004, Vol. I, pp. 505–508.
[123] J. Ma and L. Deng. “Effcient decoding strategies for conversational speech recognition
using a constrained nonlinear state–space model for vocal–tract–resonance dynamics,”
IEEE Trans. Speech Audio Process., Vol. 11, No. 6, 2003, pp. 590–602.
[124] L. Deng, D. Yu, and A. Acero. “A long-contextual-span model of resonance dynamics
for speech recognition: Parameter learning and recognizer evaluation,” Proceedings of
the IEEE Workshop on Automatic Speech Recognition and Understanding, Puerto Rico,
Nov. 27 – Dec 1, 2005, pp. 1–6 (CDROM).
[125] M. Pitermann. “Effect of speaking rate and contrastive stress on formant dynamics and
vowel perception,” J. Acoust. Soc. Am., Vol. 107, 2000, pp. 3425–3437.
[126] L. Deng, L. Lee, H. Attias, and A. Acero. “A structured speech model with continuous
hidden dynamics and prediction-residual training for tracking vocal tract resonances,”
IEEE Proceedings of the ICASSP, Montreal, Canada, 2004, pp. 557–560.
[127] J. Glass. “A probabilistic framework for segment-based speech recognition,” Comput.
Speech Language, Vol. 17, No. 2/3, pp. 137–152.
[128] A. Oppenheim and D. Johnson. “Discrete representation of signals,” Proc. IEEE,
Vol. 60, No. 6, 1972, pp. 681–691.
P1: IML/FFX P2: IML
MOBK024-BIB MOBK024-LiDeng.cls May 16, 2006 17:39

104
P1: IML/FFX P2: IML
MOBK024-AUTH MOBK024-LiDeng.cls May 30, 2006 12:33

105

About the Author

Li Deng received the B.Sc. degree in 1982 from the University of Science and Technology of
China, Hefei, M.Sc. in 1984 and Ph.D. degree in 1986 from the University of Wisconsin –
Madison.
Currently, he is a Principal Researcher at Microsoft Research, Redmond, Washington,
and an Affiliate Professor of Electrical Engineering at the University of Washington, Seattle
(since 1999). Previously, he worked at INRS-Telecommunications, Montreal, Canada (1986–
1989), and served as a tenured Professor of Electrical and Computer Engineering at University of
Waterloo, Ontario, Canada (1989–1999), where he taught a wide range of electrical engineering
courses including signal and speech processing, digital and analog communications, numerical
methods, probability theory and statistics. He conducted sabbatical research at Laboratory
for Computer Science at Massachusetts Institute of Technology (1992–1993) and at ATR
Interpreting Telecommunications Research Laboratories, Kyoto, Japan (1997–1998). He has
published over 200 technical papers and book chapters, and is inventor and co-inventor of
numerous U.S. and international patents. He co-authored the book “Speech Processing—A
Dynamic and Optimization-Oriented Approach” (2003, Marcel Dekker Publishers, New York),
and has given keynotes, tutorials and other invited lectures worldwide. He served on Education
Committee and Speech Processing Technical Committee of the IEEE Signal Processing Society
(1996–2000), and was Associate Editor for IEEE Transactions on Speech and Audio Processing
(2002–2005). He currently serves on Multimedia Signal Processing Technical Committee, and
on the editorial boards of IEEE Signal Processing Magazine and of EURASIP Journal on Audio,
Speech, and Music Processing. He is a Technical Chair of IEEE International Conference
on Acoustics, Speech, and Signal Processing (ICASSP 2004) and General Chair of IEEE
Workshop on Multimedia Signal Processing (MMSP 2006). He is Fellow of the Acoustical
Society of America and Fellow of the IEEE.
P1: IML/FFX P2: IML
MOBK024-AUTH MOBK024-LiDeng.cls May 30, 2006 12:33

106