台語語音合成技術之研究

Abstract

本論文利用基於隱藏式馬可夫模型實作台語語音合成系統，並合併國語字典與台語字典建立出一國語對台語翻譯字典，我們利用此國對台翻譯字典將中文文章翻譯為台語文章，再將翻譯後的文章，輸入至台語語音合成系統，實作出一中文文章輸入台語語音合成輸出系統。 最後考慮到台語與中文語法結構上的相似性，我們訓練一中文韻律階層式模型，並透過此模型之訓練結果，來對台語語料庫分析聲調影響台語的音節長度、音節能量及台語停頓分佈情形等特性。

In this thesis, we propose a Taiwanese speech synthesis system based on Hidden Markov Model and combine Mandarin dictionary and Taiwanese dictionary to build up a Mandarin-to-Taiwanese translation dictionary. With this dictionary, Mandarin articles can be translated into Taiwanese articles as input for Taiwanese speech synthesis system. Hence, a Taiwanese speech synthesis system with Mandarin input article is implemented. Consider the similarities on the grammar structure of Mandarin and Taiwanese, we train a Mandarin Hierarchical prosody model and use the training result to analysis the Taiwanese tone effect on Taiwanese syllable duration,syllable energy and pause distribution of Taiwanese speech database.