漢語多種腔調語速相依韻律模型之建立與其在語音合成之應用

Abstract

本論文應用現有的語速相依階層式韻律模型(SR-HPM)來探討漢語腔調的模式，首先由204位語者所產生的語料建立一個多語者SR-HPM，接著使用調適訓練技術，將多語者SR-HPM當作樣本模型來產生四個腔調的SR-HPMs，藉由分析此四個腔調的SR-HPMs我們可以觀察到個別腔調的許多韻律發音的特性，這些實驗結果與我們現有的語言學知識相符。最後實作完成個人化TTS系統，讓合成語音的韻律符合個人所屬的腔調，由主客觀評測證實這些個人化TTS系統有很好的效能。

This thesis discusses the accent modeling of multi-speaker Mandarin speech based on the existing speaker-dependnet hierarchical prosodic model (SR-HPM). It first constructs a compact multi-speaker SR-HPM using a speech corpus produced by 204 speakers with different accents. It then adopts the adaptative training technique to construct four accent-dependent SR-HPMs with the multi-speaker SR-HPM as the reference model. Through analyzing these four models, many distinct prosody pronunciation features for each accent of Mandarin speech can be found. These observations conform to our prior linguistic knowledge. An application of using these accent-dependnet SR-HPMs to construct personalized TTS systems with their own accent is realized. Both objective and subjective tests confirmed the high performances of these TTS systems