Pronunciation Variation Analysis and Modeling for Mandarin Chinese for Improved Speech Recognition

• Main Authors: Ming-Yi Tsai, 蔡明怡
• Other Authors: Lin-Shan Lee
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/64768831400098720285

博士 === 國立臺灣大學 === 電信工程學研究所 === 94 === This thesis consists of two parts, one on pronunciation variation analysis and the other on pronunciation modeling, both for Mandarin Chinese. In the first part of the thesis, the pronunciation variation for Mandarin Chinese was extensively analyzed in a quantitative way. Various statistical methods were used for the analysis, including the proposed acoustic and phonemic distances in addition to pronunciation entropy and phonological rules. The pronunciation entropy were used to analyze the dependency of pronunciation variation at different linguistic levels on various contextual conditions, different speaking rates and different occurring frequencies. On the other hand, the proposed framework based on the acoustic/phonemic distances was used for analyzing the acoustic and phonemic confusion between Initial/Finals or phonemes. Furthermore, the probabilistic phonological rules were derived automatically from speech data to analyze the phonological transformation in various context conditions. All these analyses were carried out on planned (LDC HUB-4NE) and spontaneous (LDC CALLHOME) Mandarin Chinese speech corpora. On the other hand, multiple-pronunciation dictionaries have been found to be useful in pronunciation modeling for speech recognition. However, the extra pronunciation variants added in the dictionary inevitably increase the confusion among different words during recognition, and consequently limit the achievable improvements in the recognition performance. The second part of this thesis therefore further proposed a three-stage framework for Mandarin Chinese to construct automatically the multiple-pronunciation dictionary while reducing the possible confusion caused. The proposed framework includes pronunciation generation (Stage 1), ranking (Stage 2) and pruning (Stage 3). New measures of confusability for multiple-pronunciation dictionaries were developed and shown to have a very strong correlation with the recognition performance. With the proposed framework, it was shown that the confusability as measured can be reduced and recognition performance improved stage by stage. To further reduce the possible confusion during recognition, it was then proposed that the pronunciation probabilities in the multiple-pronunciation dictionaries can be re-estimated within a proposed rapid discriminative training framework using simulated recognition errors based on a Speech Production/Recognition Model. The experimental results show that the recognition performance can be improved over the training iterations. These findings were verified by a series of experiments performed on planned (LDC HUB-4NE) and spontaneous (LDC CALLHOME) Mandarin Chinese speech corpora.