Research On Spectral Modeling And Parameter Generation In Statistical Parametric Speech

With the development of statistical modeling techniques for speech signals and computing ability of computers, statistical parametric speech synthesis methods have made significant progress in recent years. One representative approach of these methods is Hidden Markov Model (HMM) based parametric synthesis, which has become a mainstream speech synthesis because of its advantages, such as high smoothness, automatic system construction, small system footprint, and so on. However, the quality and the naturalness of the synthetic speech are degraded due to inadequacy of acoustic in the synthesizer and the over-smoothing effect of parameter generation.Therefore, this thesis is about the research on spectral modeling and parameter generation in HMM-based statistical parametric speech synthesis. Firstly, the author proposes a modeling method that integrates global variance of log power spectrum into minimum generation error training, in order to achieve better naturalness of synthetic speech without increasing the calculation at the synthesis stage. Secondly, a new RBM-based spectral modeling and parameter generation methods is proposed in this thesis, which can avoid the loss of the model precision in Gaussian approximation method. Finally, the author integrates RBM modeling with parameter generation considering GV. This method can improve the naturalness of synthetic speech further.The whole thesis is organized as follows:Chapter1is the introduction. It briefly introduces statistical parametric speech synthesis methods and reviews the history and state-of-art of this research area.Chapter2introduces a spectral modeling method on linear spectral pairs, which integrates global variance of log power spectrum into MGE training. This proposed method can achieve better naturalness of synthetic speech than the conventional MGE model training without loss of efficiency at synthesis time when LSPs are used as spectral features.Chapter3introduces research on RBM-based spectral modeling and parameter generation methods. A method of estimating mode by Gaussian approximation is designed which uses Gibbs sampling, and it can achieve similar results with Gaussian approximation method. Furthermore, the thesis proposes a RBM-HMM based parameter generation method with the constraints of dynamic features, which can achieve higher quality synthetic speech than the method using Gaussian approximation parameter generation.Chapter4introduces a parameter generation method integrating parameter generation considering GV with RBM modeling. Two methods are proposed and compared in this chapter. The results show that the proposed method can improve the naturalness of synthetic speech compared with RBM modeling and the parameter generation considering GV.Chapter5concludes the whole dissertation.