台語TTS系統之改進

Abstract

在本論文中，我們對交通大學語音實驗室以前開發的台語文句翻語音系統做了如下的改進，首先，為了避免台文表示法紊亂所造成文句分析上的困難，建置了一個國語對照台文拼音的辭典，使得此系統具備了處理國語文章的『翻譯』能力；其次，為了兼顧系統日後的擴充性對斷詞規則作了改造，並加入數詞的構詞法則；接著，利用相關函數的方法求出訓練語料的基頻標記，藉著這些基頻標記抽取更準確、有效的韻律訊息，並重新改造RNN韻律訊息產生器之架構，調整輸入參數，以改進RNN韻律訊息產生器之訓練，產生更接近自然語音流暢的韻律訊息；最後，將上述的作法和之前的系統介面整合。

In this thesis, an improvement of the TTS system, developed previously in the Speech Processing Lab of NCTU, on the following aspects is presented. Firstly, a “Chinese-to-Taiwanese” lexicon is built to solve the out-of-vocabulary problem of text analysis using Taiwanese lexicons only. It also makes the system possess the capability of processing Chinese text. Secondly, several rules of constructing compound words are added. Thirdly, a correlation-based pitch mark detection is employed to provide more accurate pitch information for the training of the RNN prosody synthesizer. Fourthly, the structure of RNN prosody synthesizer is changed. It uses one RNN with different input features for the generation of each type of prosodic information. Lastly, a new prototype Taiwanese TTS system is implemented on the PC with Windows 98/NT.