published_or_final_version / English Studies and Comparative Literature / Master / Master of Philosophy

(Uncorrected OCR) Abstract of thesis entitled Speaker- Independent Recognition of Putonghua Finals submitted by CHAN, Chit Man for the degree of Doctor of Philosophy at the University of Hong Kong � In December 1987 ABSTRACT A detailed study had been performed to address the problem of speaker-independent recognition of Putonghua (Mandarin) finals. The study included 35 Putonghua finals, 16 of which having trailing nasals. They were spoken by 51 speakers: 38 females, 13 males, in 5 different tones for two times. The sample was spectrally analyzed by a bank of 18 nonoverlapping critical-band filters. Three data reduction techniques: Karhunen-Loeve Transformation (KLT) , Discrete Cosine Transformation (OCT) and Stepwise Discriminant Analysis (SDA) , were comparat i vely studied for their feature representation capability. The results indicated that KLT was superior to both OCT and SDA. Furthermore, the theoretic equivalence of OCT to KLT was found to be valid only with 5 or more feature dimensions used in computation. On the other hand, the results also showed that the Hahalanobis and a proposed modified Mahalanobis distance both gave a better measurement of performance than the other distances tested, which included the City Block, Euclidean, Minkowski, and Chebyshev. .,. In the second Part of the study, the Hidden Markov Modelling (HMM) technique was investigated. Three classification methods: Phonemic Labell ing (PL), Vector Quantization (VQ) and a proposed Hybrid Symbol (HS) generation, were studied for use with HMM. Whilst PL was found to be simple and efficient, its performance was not as good as VQ. However, the time taken by VQ was excessive, especially in training. The results with the HS method showed that it .could successfully merge the speed advantage of PL and the better discriminatory power of VQ. An approximately 80% saving in the quantizer training time could be achieved with only a marginal loss in performance. At the same time, it Abs-l Abstract was also found that allowing skipping of states in a Left-to-Right model (LRM) could lead to a negative effect on overall recognition. As an indication of performance, the recognition rate of the simulated system was 81.3%, 95.0% and 98.0% with the best I, 2, and 3 candidates included, respectively, using a 256-level VQ and a 6-state, no-skip LRM on a sample of 8,400 finals from 48 speakers. The specific rates on non-nasal finals achieved even 96% - 98% using the best candidate alone . .. ," Abs-2 / abstract / toc / Electrical and Electronic Engineering / Doctoral / Doctor of Philosophy

Although tone plays an important role in Chinese speech perception, many issues regarding tone processing remained debatable. Among these issues, whether tone and segment are integrally or independently processed is the one that motivated this study. To investigate the integration vs. independence view of tone and segment (i.e., rime in this study) processing during Mandarin sentence comprehension, the present experiment adopted the violation paradigm and event- related potential (ERP) recording. Participants heard a series of sentences and judged whether each sentence made sense or not after its offset. The sentences included congruous sentences that embodied the original word (e.g., "观众"/guan1-zhong4/; "audience") and incongruous sentences that were created by mismatching tone (e.g., "观肿"/guan1-zhong3/), rime (e.g., " 观赚"/guan1-zhuan4/) or tone-plus-rime (e.g., "观转"/guan1-zhuan3/) of the second syllable of the original words in the congruous sentences. Larger N400 (250-400 ms) and P600 (500-700 ms) were evoked by the incongruous sentences than congruous sentences. Among the incongruous sentences, the N400 elicited by double violation (i.e., tone-plus-rime violation) was larger than rime violation, which was in turn larger than tone violation. The P600 evoked by tone violation was larger than rime violation in 500-600 ms, but they were comparable in 600-700 ms. In addition, the P600 evoked by tone and rime violation were both larger than double violation. The different ERP effects among the three violation conditions supported the independence view of tone and rime processing. Based on the results, a dynamic model of spoken word processing was proposed, in which the functional dissociation of tone and segment across different stages was taken into consideration.

by Choy Chi Yan. / Thesis (M.Phil.)--Chinese University of Hong Kong, 1999. / Includes bibliographical references (leaves 111-115). / Text in English; abstract also in Chinese. / by Choy Chi Yan. / ABSTRACT --- p.I / ACKNOWLEDGMENTS --- p.III / TABLE OF CONTENTS --- p.IV / LIST OF FIGURES --- p.VII / LIST OF TABLES --- p.VIII / Chapter 1. --- INTRODUCTION --- p.1 / Chapter 1.1 --- Speech Recognition --- p.1 / Chapter 1.2 --- Development of Speech Recognisers --- p.4 / Chapter 1.3 --- Speech Recognition for Chinese Language --- p.5 / Chapter 1.4 --- Objectives of the thesis --- p.6 / Chapter 1.5 --- Thesis Structure --- p.7 / Chapter 2. --- PHONOLOGICAL AND ACOUSTICAL PROPERTIES OF MANDARIN CHINESE --- p.9 / Chapter 2.1 --- Characteristics of Mandarin Chinese --- p.9 / Chapter 2.1.1 --- Syllabic Structures --- p.10 / Chapter 2.1.2 --- Lexical Tones --- p.11 / Chapter 2.2 --- Basic Phonetic Units for Mandarin Chinese --- p.14 / Chapter 2.2.1 --- Tonal Syllables and Base Syllables --- p.14 / Chapter 2.2.2 --- Initial-Finals --- p.14 / Chapter 2.2.3 --- Phones --- p.16 / Chapter 2.2.4 --- Preme-Core-Finals and Preme-Tonemes --- p.17 / Chapter 2.2.5 --- Summary-The phonological hierarchy of Mandarin Syllables --- p.19 / Chapter 3. --- HIDDEN MARKOV MODELS --- p.20 / Chapter 3.1 --- Introduction --- p.20 / Chapter 3.1.1 --- Speech Data --- p.20 / Chapter 3.1.2 --- Fundamental of HMMs --- p.21 / Chapter 3.2 --- Using Hidden Markov Models for Speech Recognition --- p.22 / Chapter 3.2.1 --- Likelihood of the state sequence of speech observations --- p.22 / Chapter 3.2.2 --- The Recognition Problem --- p.24 / Chapter 3.3 --- Output Probability Distributions --- p.25 / Chapter 3.4 --- Model Training --- p.26 / Chapter 3.4.1 --- State Sequence Estimation --- p.26 / Chapter 3.4.2 --- Gaussian Mixture Models --- p.29 / Chapter 3.4.3 --- Parameter Estimation --- p.30 / Chapter 3.5 --- Speech Recognition and Viterbi Decoding --- p.31 / Chapter 3.6 --- Summary --- p.32 / Chapter 4. --- LARGE VOCABULARY CONTINUOUS SPEECH RECOGNITION FOR MANDARIN CHINESE --- p.33 / Chapter 4.1 --- Introduction --- p.33 / Chapter 4.2 --- Large Vocabulary Mandarin Chinese Recognition System --- p.34 / Chapter 4.2.1 --- Overall Architecture for the Speech Recogniser --- p.34 / Chapter 4.2.2 --- Signal Representation and Features --- p.36 / Chapter 4.2.3 --- Subword Unit Models Based on HMMs --- p.39 / Chapter 4.2.4 --- Training of Subword Units --- p.42 / Chapter 4.2.5 --- Language Model (LM) --- p.43 / Chapter 4.2.6 --- "Transcriptions, Word Networks and Dictionaries for LVCSR System" --- p.44 / Chapter 4.2.7 --- Viterbi Decoding --- p.47 / Chapter 4.2.8 --- Performance Analysis --- p.48 / Chapter 4.3 --- Experiments --- p.48 / Chapter 4.3.1 --- Tasks --- p.48 / Chapter 4.3.2 --- Speech Database --- p.49 / Chapter 4.3.3 --- Baseline Experimental Results --- p.51 / Chapter 4.4 --- Context Dependency in Speech --- p.52 / Chapter 4.4.1 --- Introduction --- p.52 / Chapter 4.4.2 --- Context Dependent Phonetic Models --- p.53 / Chapter 4.4.3 --- Word Boundaries and Word network for context-dependent HMMs --- p.54 / Chapter 4.4.4 --- Recognition Results Using Cross-Syllable Context-Dependent Units --- p.56 / Chapter 4.5 --- Tree-Based Clustering --- p.58 / Chapter 4.5.1 --- Introduction --- p.58 / Chapter 4.5.2 --- Decision Tree Based Clustering --- p.59 / Chapter 4.5.3 --- The Question Sets --- p.61 / Chapter 4.5.4 --- Convergence Condition --- p.61 / Chapter 4.4.5 --- The Final Results --- p.63 / Chapter 4.6 --- Conclusions --- p.65 / Chapter 5. --- APPLICATION1 ISOLATED WORD RECOGNITION FOR MANDARIN CHINESE --- p.67 / Chapter 5.1 --- Introduction --- p.67 / Chapter 5.2 --- Isolated Word Recogniser --- p.68 / Chapter 5.2.1 --- System Description --- p.68 / Chapter 5.2.2 --- Experimental Results --- p.70 / Chapter 5.3 --- Discussions and Conclusions --- p.71 / Chapter 6. --- APPLICATION2 SUBWORD UNITS FOR A MANDARIN KEYWORD SPOTTING SYSTEM --- p.74 / Chapter 6.1 --- INTRODUCTION --- p.74 / Chapter 6.2 --- RECOGNITION SYSTEM DESCRIPTION --- p.75 / Chapter 6.2.1 --- Overall Architecture and Recognition Network for the keyword Spotters --- p.75 / Chapter 6.2.2 --- Signal Representation and Features --- p.76 / Chapter 6.2.3 --- Keyword Models --- p.76 / Chapter 6.2.4 --- Filler Models --- p.77 / Chapter 6.2.5 --- Language Modeling and Search --- p.78 / Chapter 6.3 --- EXPERIMENTS --- p.78 / Chapter 6.3.1 --- Tasks --- p.78 / Chapter 6.3.2 --- Speech Database --- p.79 / Chapter 6.3.3 --- Performance Measures --- p.80 / Chapter 6.3.4 --- Details of Different Word-spotters --- p.80 / Chapter 6.3.5 --- General Filler Models --- p.81 / Chapter 6.4 --- EXPERIMENTAL RESULTS --- p.83 / Chapter 6.5 --- CONCLUSIONS --- p.84 / Chapter 7. --- CONCLUSIONS --- p.87 / Chapter 7.1 --- Review of the Work --- p.87 / Chapter 7.1.1 --- Large Vocabulary Continuous Speech Recognition for Mandarin Chinese --- p.87 / Chapter 7.1.2 --- Isolated Word Recognition for a Stock Inquiry Application --- p.88 / Chapter 7.1.3 --- Keyword Spotting for Mandarin Chinese --- p.89 / Chapter 7.2 --- Suggestions for Further Work --- p.89 / Chapter 7.3 --- Conclusion --- p.91 / APPENDIX --- p.92 / BIBLIOGRAPHY --- p.111