自發性對話語音辨識之初步研究

Abstract

在本論文中，我們建立一個自發性中文對話語音辨識基本系統架構，探討中文語音及自發性語料的特殊語音現象，如感嘆語(particles)、語音發音變異(uncertain sounds)、非語音聲音 (paralinguistic sounds)等，之聲學模型建立方法，使用中研院提供的八個雙人對話語料做實驗，獲得之音節辨識率為43.33%。為使辨識系統更為完善，我們加入語言模型，並以語言調適的技術，使之更為優化，最後音節辨識率達到53.93%，較基本系統提升了10.6%。

In the thesis, a basic spontaneous Mandarin speech recognition system is established. The study focuses on the acoustic modeling for 411 Mandarin base-syllables as well as some special phenomena of spontaneous speech such as particles, uncertain sounds, and paralinguistic phenomena. Performance of the proposed system was examined by simulations using a Mandarin dialogue speech database called MCDC (Mandarin Conversational Dialogue Corpus). A syllable accuracy rate of 43.33% was obtained. By adding a bi-gram language model with proper adaptation, the syllable accuracy rate increased to 53.93% which was 10.6% better than the baseline system.