台語語音辨識與文字處理之研究

Abstract

本論文探討台語語音處理的兩個問題：語音辨識與文句分析，在台語語音辨識方面，我們建立聲韻母次音節隱藏式馬可夫模型，作連續語音音節辨識，在多語者的情況下，男女生之音節辨識率分別為42.67%及47.33%；在台語文句分析方面，我們使用詞典依長詞優先之原則進行斷詞，對於台語文字表示法不統一而引起的斷詞混淆問題，我們藉建立音節與字的對應表，將詞典之詞展開，來改善斷詞之正確率。 關鍵詞：台語語音辨識、隱藏式馬可夫模型、台語文句分析、斷詞

In this thesis, two tasks of Taiwanese language processing are studied. First, the task of Taiwanese speech recognition is exploited. A set of initial and final sub-syllable hidden Markov models (HMMs) is constructed for continuous base-syllable recognition. Syllable accuracy rates of 47.33% and 42.67% were obtained for multi-speaker female and male speech recognition, respectively. Then, the task of Taiwanese text analysis is studied. The problem of word tagging ambiguity due to the non-standardization of Taiwanese written form is exploited. A method to expand the representations of words in the lexicon by using a syllable-to-character mapping table is proposed. Experimental results confirmed the effectiveness of the proposed method on improving the performance of word tagging. Keywords: Taiwanese speech recognition, hidden Markov model, Taiwanese text analysis, word tagging