With the development of speech synthesis, the unit selection speech synthesis system based on statistical parametric models has caught a great of researchers’ attention recently. Hidden Markov Model based hybrid unit selection speech synthesis system combines the advantages of statistical parametric speech synthesis system and unit selection speech synthesis system, and the quality of synthesized speech is improved. HUS system is still in the initial stage and has lots of disadvantages. So far, there is no hybrid unit selection system could meet the demand of the market totally. In order to improve the quality of synthesized speech, performance of prosody and the system speed, research has been carried out in pre-selection, unit selection and the cost calculation method in this paper. The main research work and results are as follows in this paper: The first chapter is the introduction. We reviewed the research history of speech synthesis. Classical speech synthesis methods are introduced in detail and the research objectives are presented here. The second chapter introduced the related concepts and framework of speech concatenation system. In order to improve the prosody performance in the speech concatenation system, context features based unit selection system is put forward in this chapter. Without using traditional machine learning methods to predict the acoustic parameters, context features is considered. The context features after the text analysis is used to guiding the units selection. Linear regression and decision tree based M5P algorithm is applied to calculate the target cost. A hierarchical pre-selection method is proposed to improve the speed of the system. Especially the duration predicted model added in the hierarchical pre-selection ensures the stability in the duration of the selected units. Experiments show that context features based unit selection system has been greatly improved in the naturalness of the synthesized speech. The third chapter focuses on the HMM based hybrid unit selection system and a novel system is presented in this chapter. Firstly, HMM based hybrid unit selection is introduced in detail including the acoustic modeling and system framework. Several typical HMM-based hybrid unit selection systems are introduced in this chapter. On this basis, a data driven based hybrid unit selection system is proposed. This approach combines the results of the previous chapter. Context feature based multiple linear regression ...
修改评论