透過您的圖書館登入
IP:3.144.77.71
  • 學位論文

基於類神經網路之中文語音停頓預估

Neural Network-based Break Prediction for Mandarin Speech

指導教授 : 陳信宏

摘要


本論文針對語音合成中的停頓標記議題作改進,利用以聲學參數以及語言參數輔助下產生之停頓標記作為研究之預估目標。由於在文字轉語音系統中,並無聲學參數輔助,僅能透過文字中的語言參數預估停頓標記。本研究基於人工類神經網路之架構透過語言參數預估停頓,並且試以不同網路架構改善其預估效能。我們另外加入語法樹中語詞組相關的語言參數,透過豐富語言參數的方式來取得更佳的停頓標記,最後再利用類神經網路產生合成所需的停頓時長。實驗結果顯示加入語法樹的部分資訊即可在停頓標記方面有所幫助,而在以類神經網路預估停頓時長也有部分改善。

並列摘要


The purpose of this thesis is to improve the break tags prediction which is generated form acoustic feature and linguistic feature for Mandarin speech synthesis. However there are no acoustic features in TTS synthesis system, linguistic features extracted from context data are the only imformation. This research utilize different Artitficial Neural Network model to predict break tags. In order to improve the break prediction, more linguistic feature descrbing for syntactic information are added. Furthemore, we predict pause duration based on Neural Network. The experiment result showed that partial imformation of syntax tree can improve break prediction actually, and the performance of the pause prediction is also improved.

參考文獻


[1] Y.-Q. Shao, Y.-Z. Zhao, J.-Q. Han, and T. Liu, “Using different models to label
[2] C. W. Wightman and M. Ostendorf, “Automatic labeling of prosodic patterns,”
IEEE Trans. Speech Audio Process. Vol. 2, pp. 469–481, 1994.
[3] X. Sun and T. H. Applebaum, “Intonational phrase break prediction using
[5] A. W. Black and P. Taylor, “Assigning phrase breaks from part-of-speech

延伸閱讀