Table Of ContentA Syllable, Articulatory-Feature, and Stress-Accent Model of Speech
Recognition
by
Shuangyu Chang
B.S.E. (University of Michigan, Ann Arbor) 1997
A dissertation submitted in partial satisfaction of the
requirements for the degree of
Doctor of Philosophy
in
Computer Science
in the
GRADUATE DIVISION
of the
UNIVERSITY of CALIFORNIA, BERKELEY
Committee in charge:
Professor Nelson Morgan, Cochair
Dr. Lokendra Shastri, Cochair
Dr. Steven Greenberg
Professor Edwin R. Lewis
Professor David L. Wessel
Professor Lot(cid:12) A. Zadeh
Fall 2002
The dissertation of Shuangyu Chang is approved:
Cochair Date
Cochair Date
Date
Date
Date
Date
University of California, Berkeley
Fall 2002
A Syllable, Articulatory-Feature, and Stress-Accent Model of Speech
Recognition
Copyright (cid:13)c Fall 2002
by
Shuangyu Chang
1
Abstract
A Syllable, Articulatory-Feature, and Stress-Accent Model of Speech Recognition
by
Shuangyu Chang
Doctor of Philosophy in Computer Science
University of California, Berkeley
Professor Nelson Morgan, Dr. Lokendra Shastri, Cochairs
Current-generation automatic speech recognition (ASR) systems assume that
wordsarereadilydecomposableintoconstituentphoneticcomponents(\phonemes").
A detailed linguistic dissection of state-of-the-art speech recognition systems indi-
cates that the conventional phonemic \beads-on-a-string" approach is of limited
utility, particularly with respect to informal, conversational material. The study
shows that there is a signi(cid:12)cant gap between the observed data and the pronunci-
ation models of current ASR systems. It also shows that many important factors
a(cid:11)ecting recognition performance are not modeled explicitly in these systems.
Motivated by these (cid:12)ndings, thisdissertation analyzes spontaneous speech with
respect to three important, but often neglected, components of speech (at least
with respect to English ASR). These components are articulatory-acoustic fea-
tures (AFs), the syllable and stress accent. Analysis results provide evidence for
an alternative approach of speech modeling, one in which the syllable assumes pre-
2
eminent status and is melded to the lower as well as the higher tiers of linguistic
representation through the incorporation of prosodic information such as stress
accent. Using concrete examples and statistics from spontaneous speech material
it is shown that there exists a systematic relationship between the realization of
AFs and stress accent in conjunction with syllable position. This relationship can
be used to provide an accurate and parsimonious characterization of pronunciation
variation in spontaneous speech. An approach to automatically extract AFs from
the acoustic signal is also developed, as is a system for the automatic stress-accent
labeling of spontaneous speech.
Based on the results of these studies a syllable-centric, multi-tier model of
speech recognition is proposed. The model explicitly relates AFs, phonetic seg-
ments and syllable constituents to a framework for lexical representation, and in-
corporates stress-accent information into recognition. A test-bed implementation
of the model is developed using a fuzzy-based approach for combining evidence
from various AF sources and a pronunciation-variation modeling technique using
AF-variation statistics extracted from data. Experiments on a limited-vocabulary
speech recognition task using both automatically derived and fabricated data
demonstrate the advantage of incorporating AF and stress-accent modeling within
the syllable-centric, multi-tier framework, particularly with respect to pronuncia-
tionvariationinspontaneousspeech.
Professor Nelson Morgan
Dissertation Committee Cochair
Dr. Lokendra Shastri
Dissertation Committee Cochair
i
To Jiangxin
ii
Contents
List of Figures vii
List of Tables xvi
1 Introduction 1
1.1 The Conventional Model of Speech Recognition . . . . . . . . . . . 1
1.2 Finding Alternatives . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Linguistic Dissection of LVCSR Systems 12
2.1 Background Information . . . . . . . . . . . . . . . . . . . . . . . . 13
2.1.1 Corpus Material . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.1.2 Participating Systems . . . . . . . . . . . . . . . . . . . . . 16
2.2 Analysis Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Word and Phone Error Patterns . . . . . . . . . . . . . . . . 17
2.2.2 Syllable Structure and Syllable Position . . . . . . . . . . . . 25
2.2.3 Articulatory-acoustic Features and Syllable Position . . . . . 28
2.2.4 Prosodic Stress Accent and Word Errors . . . . . . . . . . . 32
2.2.5 Speaking Rate and Word Errors . . . . . . . . . . . . . . . . 35
iii
2.2.6 Pronunciation Variation and Word Errors . . . . . . . . . . 37
2.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 Articulatory-acoustic Features 43
3.1 Background and Previous Work . . . . . . . . . . . . . . . . . . . . 44
3.2 Automatic Extraction of Articulatory-acoustic Features . . . . . . . 48
3.2.1 System Description . . . . . . . . . . . . . . . . . . . . . . . 48
3.2.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.3 Extension to Automatic Phonetic Labeling . . . . . . . . . . 56
3.3 Manner-speci(cid:12)c Training and the \Elitist" Approach . . . . . . . . 61
3.3.1 AF Classi(cid:12)cation on the NTIMIT Corpus . . . . . . . . . . 61
3.3.2 An \Elitist" Approach . . . . . . . . . . . . . . . . . . . . . 62
3.3.3 Manner-Speci(cid:12)c Training . . . . . . . . . . . . . . . . . . . . 67
3.4 Cross-linguistic Transfer of AFs . . . . . . . . . . . . . . . . . . . . 69
3.5 Robustness of AFs . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.5.1 Corpus Material with Noise . . . . . . . . . . . . . . . . . . 76
3.5.2 Experimental Results . . . . . . . . . . . . . . . . . . . . . . 76
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4 Speech Processing at the Syllable Level 81
4.1 What is a Syllable? . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.2 The Stability and Importance of the Syllable in Speech Perception . 84
4.2.1 Stability of Syllables in Speech Corpora . . . . . . . . . . . . 85
4.2.2 Acoustic-based Syllable Detection and Segmentation . . . . 85
iv
4.2.3 Signi(cid:12)cance of Syllable Duration . . . . . . . . . . . . . . . . 87
4.2.4 Syllables and Words . . . . . . . . . . . . . . . . . . . . . . 89
4.3 Pronunciation Variation, Prosody and the Syllable . . . . . . . . . . 89
4.4 Articulatory-acoustic Features and the Syllable . . . . . . . . . . . 93
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5 Stress Accent in Spontaneous American English 104
5.1 Stress Accent in Spontaneous American English . . . . . . . . . . . 105
5.1.1 The Perceptual Basis of Stress Accent . . . . . . . . . . . . 106
5.1.2 Vocalic Identity and Stress Accent . . . . . . . . . . . . . . 107
5.2 Stress Accent and Pronunciation Variation . . . . . . . . . . . . . . 111
5.2.1 Pronunciations of \That" { Revisited . . . . . . . . . . . . . 112
5.2.2 Impact of Stress Accent by Syllable Position . . . . . . . . . 114
5.3 Automatic Stress-Accent Labeling of Spontaneous Speech . . . . . . 131
5.3.1 System Description . . . . . . . . . . . . . . . . . . . . . . . 132
5.3.2 Experiments on the Switchboard Corpus . . . . . . . . . . . 134
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
6 A Multi-tier Model of Speech Recognition 140
6.1 Model Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.2 Questions Regarding the Multi-tier Model . . . . . . . . . . . . . . 145
6.3 Test-bed System Implementation . . . . . . . . . . . . . . . . . . . 146
6.3.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.3.2 AF Classi(cid:12)cation and Segmentation . . . . . . . . . . . . . . 152
v
6.3.3 Stress-accent Estimation . . . . . . . . . . . . . . . . . . . . 154
6.3.4 Word Hypothesis Evaluation . . . . . . . . . . . . . . . . . . 155
6.3.5 Cross-AF-dimension Syllable-score Combination . . . . . . . 158
6.3.6 Within-syllable Single-AF-dimension Matching . . . . . . . . 164
6.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7 Multi-tier Recognition { Experiments and Analysis 171
7.1 Experimental Conditions . . . . . . . . . . . . . . . . . . . . . . . . 172
7.2 Overall System Performance . . . . . . . . . . . . . . . . . . . . . . 174
7.3 Testing the Contribution of Stress Accent . . . . . . . . . . . . . . . 181
7.4 Testing Pronunciation Modeling . . . . . . . . . . . . . . . . . . . . 182
7.5 Testing the Contribution of Syllable Position . . . . . . . . . . . . . 185
7.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
8 Conclusions and Future Work 190
8.1 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . 191
8.1.1 Linguistic Dissection of LVCSR Systems . . . . . . . . . . . 191
8.1.2 Detailed Analysis of the Elements of Speech . . . . . . . . . 192
8.1.3 An Alternative Model of Speech . . . . . . . . . . . . . . . . 195
8.2 Future Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
8.2.1 Incorporation into Conventional Systems . . . . . . . . . . . 198
8.2.2 Further Analysis and Experiments . . . . . . . . . . . . . . . 200
8.2.3 An Improved Framework and Implementation . . . . . . . . 202
8.3 Coda . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203