臺灣博碩士論文加值系統

本論文為國台雙語大詞彙與連續語音辨認之研究，我們採用HMM的原理來做語音模型的訓練與辨認，採用右相聯音素模型，在臺語大辭彙辨認方面，五萬詞的辨認率可達到93.17%，兩萬詞的辨認率可達到95.98%，在國台雙語大詞彙辨認率方面，四萬詞的辨認率可達89.26%，連續語音方面的方面，使用回溯式樹狀網路架構改善了雙音節雙連文法網路的速度，並得到76.36%的辨認率。最後我們將介紹在windows平台上所發展的語音辨認即時系統。

This thesis is about a large vocabulary, continuous speech recognition system for bi-lingual Mandarin/Taiwanese (Min-nan) speech. The kernel technology used is the Hidden Markov Model which models the inside-syllabic, inside-syllable right-context-dependent phonemes.
We achieve word recognition rate 93.17% in 50k-word vocabulary, 95.98% in 20k-word vocabulary for Taiwanese speech and 89.26% in 40k-word vocabulary for bi-lingual Mandarin/Taiwanese speech ,which is uttered in isolated words by a specific speaker. In continuous speech, word recognition rate 76.36% is achieved with back-traced tree searching net.
Finally, we have a real-time recognition system with several toolkits developed on the MS-windows platform.

第一章緒論1
1.1 前言1
1.2 本論文研究的範疇1
1.3 TWBET標音系統介紹2
1.4 本論文介紹4
第二章 語音辨認完整流程與原理5
2.1 簡介5
2.2 特徵擷取 (Feature Extraction)6
2.3 隱藏式馬可夫模型6
2.4 維特比搜尋演算法7
2.5 光束搜尋(Beam Search)[2]8
2.6 表徵傳遞(Token Passing)演算法[4]9
2.7 拼音轉文字模組10
第三章 平衡詞選取方法研究11
3.1 問題定義:11
3.2 問題討論：11
3.3 兩種方法比較：17
第四章國台雙語大詞彙辨認研究19
4.1 國台雙語在聲學上的比較19
4.2 國台雙語混合聲學模型21
4.3 搜尋網路22
4.4 實驗語料結果與討論：24
第五章 台語連續語音辨認研究27
5.1 方法一 單音節網路辨認27
5.2 方法二 雙音節雙連文法網路28
5.3 方法三 回溯式樹狀網路30
5.4 實驗結果與討論33
第六章 即時語音辨認系統34
6.1 即時語音辨認系統34
6.2 標音檔案處理程式37
6.3 語音辨認工具程式38
6.4 應用實例─國台雙語醫院自動掛號系統39
6.5 本章結論40
第七章 結論41
參考文獻42
附錄一 825個台語音節列表44
附錄二407個國語音節列表48
附錄三183國台語音節交集列表50

[1] Thomas H. Cormen, Charles E. Lesiserson and Ronald L. Rivest, ”Introduction To Algorithms”
[2] Lawrence Rabiner, Biing-Hwang Juang,”Fundamentals of Speech Recognition”,Prentice Hall
[3] Xuedong Huang, Fileno Alleva, Hsiao-Wuen Hon,”The SPHINX-II Speech Recognition System: An Overview”, School of Computer Science Carnegie Mellon University, 1992
[4] S.J. Young, N.H.Russell, J.H.S Thornton, ”Token Passing : a Simple Conceptual Model for Connected Speech Recognition Systems”, Cambridge University Engineering Department, July 31, 1989
[5] 劉惠玫,”用TTS輔助台語語料之處理”,清華大學
[6] Jia-lin Shen, Hsin-min Wang, Ren-yuan Lyu and Lin-shan Lee,”Automatic selection of phonetically distributed sentence sets for speaker adaptation with application to large vocabulary Mandarin speech recognition”, Computer Speech and Language (1999) 13,p79-p98
[7]楊智祥,”國語連續語音辨認之初步研究”,長庚大學電機所
[8] Ren-yuan Lyu, Yuang-chin Chiang, Ren-jou Fang, Wen-ping Hsieh, ‘A Large-
Vocabulary Taiwanese (Min-nan) Speech Recognition System Based on Inter-syllabic Initial-Final Modeling and Lexicon-Tree Search’, ROCLING XI Conference, p.139~p.149, Aug. 1998, Hsinchu
[9] Ren-yuan Lyu, Yuang-chin Chiang, Wen-ping Hsieh, Ren-zhou Fang, Zhi-xiang, Yang,Zong-yi Lin, ‘A Large-Vocabulary Taiwanese (Min-nan) Multi-syllabic Word Recognition System Based upon Right-Context-Dependent Phones with State Clustering by Acoustic Decision Tree’, International Conference on Spoken Language Processing, Nov. 1998, Sydney, Australia
[10] Yuang-chin Chiang, Ren-zhou Fang, Wen-ping Hsieh, Ren-yuan Lyu, “A Hybrid Duration Hidden Markov Model with Application to Large Vocabulary Taiwanese (Min-nan) Speech Recognition”, International Symposium on Chinese Spoken Language Processing, Dec. 1998, Singapore
[11] Ren-Yuan Lyu, Yuang-jin Chiang, wen-ping Hsieh,”A Large-Vocabulary Taiwanese(MIN NAN) Multi-syllable Word Recognition System Based Upon Right-Context-Dependent Phones With State Clustering by Acoustic Decision Tree”,ICSLP,1998
[12] Steven Young,”The HTK Book cersion 2.1”,Cambridge University,1996
[13] Julian Odell, Dan Kershaw, Dave Ollason, Valtcho Valtchev, David Whitehouse, “The HAPI Book-A description of the HTK Application Programming Interface”