自發性對話語音音節合併現象之分析及辨識改進

Abstract

自發性對話語音是最接近人們在自然情況下的對話語音，結合語音辨識的技術可以應用在許多領域上。在本論文中，我們專注於改善自發性語音中的重要現象－音節合併現象（Syllable Contraction）的辨識，我們統計分析語料中有關音節合併現象的特性，據以對音節合併現象建立特別的聲學模型。實驗結果顯示所建立的多音節聲學模型對於發生合併現象音節之辨識效果有明顯提昇，但當音節合併現象嚴重到合併為單一音節時，仍難以正確辨識。

In this thesis, the effect of syllable contraction in spontaneous Mandarin speech recognition is exploited. Syllable contraction is a phenomenon of serious coarticulation between two consecutive syllables and is a major factor to degrade the performance of a spontaneous-speech recognizer. In the study we propose to construct separate HMM models for syllables with and without contraction. Performance of the proposed approach was examined by simulations using a Mandarin dialogue speech database called MCDC (Mandarin Conversational Dialogue Corpus). Experimental results showed that the recognition performance for contracted syllables could be greatly improved via building HMM models for contracted syllable pairs with high frequency. But the recognition is still difficult for the case of very serious contraction.