Expressive Multi-language Speech Synthesis Based on the Generation Process Model and Its Use for Automatic Speech Translation

基于基频轮廓生成过程模型（F_0模型），对多语言韵律控制进行了统一研究。我们开发了一种韵律适应方法，从平行语音语料库中学习F_0模型命令的差异，并将其应用于基线语音。实现了焦点控制、风格转换和语音转换。此外，通过逼近训练语音语料库的F_0轮廓和/或使用F_0模型生成的F_0轮廓，我们通过基于hmm的语音合成提高了合成语音的质量。此外，我们还添加了对焦控制。在此基础上，对语音翻译中话语信息和意图的传递进行了实验。

A unified study on prosody control for multi-languages was conductedbased on the generation process model of fundamental frequency contours(F_0 model). We developeda method of prosody adaptation, where differences in F_0 model commands were learned from parallelspeech corpus and were applied to baseline speech. Focus control, style conversion and voiceconversion were realized. Furthermore, by approximating F_0 contours of training speech corpusand/or generated F_0 contours using the F_0 model, we improved the quality of synthetic speech by theHMM-based speech synthesis. Also, we added focus control. Based on the above results, experiments were conducted on conveying discourse information and intentions in speech Translation.