HMM-based Trainable Speech Synthesis For Dai Language

The technique of Speech synthesis is that it aims to transform visual text information into audible speech signal by computer. In recent years, thanks to the rapid development of technology, speech synthesis technology has also made a breakthrough, it is widely used in all aspects of people’s lives.Yunnan is one of three provinces more than 10 million minority population, which inhabited by ethnic minorities. Among them, there are more than 1.2 million Dai compatriots using Dai language and mainly Distributing in in Dehong, Baoshan, Xishuangbanna area. With the economic exchanges and cultural exchanges of Dai and Han becoming more and more, the importance of the Dai language speech synthesis research is self-evident.For the purpose of developing a Dai speech synthesis system, this paper has researched on a HMM-based trainable speech synthesis for Dai language.In this paper, main work and study are as follows:1. It describes the basic principles of HMM-based speech synthesis. And then, study the implementation of Dai language trainable speech synthesis system.2. According to the feature of Xishuangbanna Dai language, complete data preparation for the Dai language speech synthesis. The main work includes four areas: construction of corpus, determination of phoneme tabulation, obtain of labeled files and design of context attribute and question collection.3. Based on HTK Toolkit and using uniform segmentation as the initial training data, HMM initial models are generated for each phone. Through multiple repeated training and phone alignment, the results of phone segmentation are automatically generated.4. Based on STRAIGHT synthesizer speech synthesis platform, use 1300 speech corpus of uniform segmentation as initial training data and then generate initial model of HMM. After tagging the corpus to be synthesized, using trained acoustic model to predict parameter. Finally, generate speech waveform by STRAIGHT synthesizers.The experimental result shows that the trained acoustic model can synthesize Dai language that intelligibility and naturalness are acceptable. The proposed synthetic scheme of Dai language and syllable automatic segmentation method are feasible.