【摘要】 语音单元边界的准确切分对基于波形拼接的语音合成系统至关重要。文章采用了两步切分方法,第一步中先由基于HMM模型的强制对齐方法得到初始的边界,在第二步中提出用基于前后音素的边界模型来修正初始边界。为解决训练数据不足的问题,提出用分类与衰退树将前后因素发音相近的边界模型进行聚类。这样可以根据训练数据的多少,动态调节边界模型的数目,以保证模型训练的可靠性。在对中文语音库的实验中,自动切分的准确度由78.7%提高到91.5%。更多还原

【Abstract】 Correct unit segmentation are, though laborsome, very crucial to the performance of a concatenation based TTS system. This paper suggests a two-step procedure for automatic unit segmentation, which coarsely segments speech data in the first step and refines segment boundaries in the secord step. A new Context-Dependent Boundary Model (CDBM) to describe the evolution across the segment boundary is proposed. To reduce manual segmentation, Classification and Regression Tree(CART) is used to structure the available data into a more efficient usage. Acoustically similar boundaries are clustered together and corresponding tied CDBM models are thus trained and used for boundary refinement during the secord step. After a series of experiments, the optimal CDBM parameters and the training conditions are found. The segmentation accuracy is raised from 78.7% to 91.5% in Mandarin syllable segmentation with about 1,000 manually segmented sentences as CDBM training data.更多还原