Research On HMM-based Voice Conversion

Recent years with the development of speech synthesis technologies, the quality and naturalness of the synthetic speech have been greatly improved. So the speech synthesis system which can synthesis various kinds of speech, with different styles or different languages of different speakers, has become the primary objectives of users. In order to meet this requirements, voice conversion by adaptation is widely used. According to a small number of target speaker's data, we adjust the original speaker's voice to similar target voice by certain methods. Therefore this thesis studies the topic of the Hidden Markov model (HMM)-based voice conversion systematically, with further research to get better results of voice conversion in different environment.Chapter 1 mainly introduces the research background, concept and method of voice conversion, including the development of speech synthesis, basic ideas, methods, advantages and disadvantages of the traditional adaptation methods.Chapter 2 gives an introduction on the whole framework of the HMM-based voice conversion. Some important parts of the speech synthesis are examined, and the most popular adaptation method-MLLR, as the basis for further improvements.Chapter 3 focuses on the EigenVoice adaptation. For better results while little data available, we train models from different speakers, and get model weights by ML rules. Experiments prove better adaptation effect compared to MLLR.Chapter 4 mainly discusses MAP rules in voice conversion. Firstly we analyze basic adaptation including MAP, SMAP, etc. Then through fully experiments in MLLR,MAP,SMAP, we can get better adaptation results, and give summary of the different adaptation methods on performance and efficiency.Chapter 5 primarily studies the cross language adaptation. Mainly introduce HMM-based adaptation from Chinese to English. Experiments confirm some effect in cross language adaptation.