一个普通话文语转换系统中的韵律模型

Prosody Model for Mandarin Text-to-Speech System

韵律模型是文语转换系统中的重要组成部分,对合成语音的自然度起着至关重要的作用。结合人工神经网络和单元选择算法,将它们分别应用于韵律模型中时长和基频曲线的生成,其中时长模型采用三层的反向传播网络,而基频模型则采用一种基于最小距离和的单元选择算法。

Prosody model is a essential part in text-to-speech system. It plays an important role in naturalness of synthesized speech. This paper integrates artificial neural networks with unit selection in prosody model, and applies them to the generation of duration and pitch. It presents a three-layer back-propagation neural network in duration model, and an algorithm based on minimizing distance summation of a whole utterance in pitch model.