Research On HMM-RBM Based Mongolian Speech Synthesis

Because computers increasingly common integrated into people’s life and work, people have an increasingly higher requirements for computer. In the aspect of human-computer interaction, people are no longer satisfied with the form of using external devices such as keyboard to input commands to computers, but rather want to communicate with computers through language directly. It makes the speech synthesis, speech recognition technology gradually become a research hotspot.The purpose of the speech synthesis technology research is that we are able to create a computer which can speak. At the present, with the strong computing power and storage capacity of the computer, speech synthesis technology has been rapid development. HMM(Hidden Markov Model) based speech synthesis method has the advantage of flexibility, portability, and synthesized speech more "humanized", which make it become one of the mainstream technology. But the speech that produced by this method has the disadvantages of over smooth, detailed loss etc. Neural network applications in the field of speech synthesis give us some idea that how to improve the naturalness of synthesis speech.Mongolian as a national language has its own characteristics, with the continued efforts of experts, Mongolian speech synthesis is constantly evolving. In this thesis, on the basis of HMM-based Mongolian speech synthesis, using Restricted Boltzmann Machine (Restricted Boltzmann Machine, RBM) for parametric modeling, use it replace single gaussian distribution to represent the state of each HMM distribution. At the same time, we use the original spectral envelope modeling directly, instead of Mel-cepstrum or lsp which was processed. Because the original spectral envelope contains more details.The experimental results show that, HMM-RBM based Mongolian speech synthesis make synthesized speech more naturalness, This result has important implications for Mongolian speech synthesis performance optimization.