Voice Conversion Research Based On Spectral Envelope And Super-segmental Prosody

Voice conversion is a technology that can change source speaker's voice character to sound like target speaker's. This technology is based on speaker verification and synthesis technology. And, it enriches and extends the two speech research fields. Voice conversion is a valuable theoretical research topic and offers many useful applications.The main work in this paper includes the following aspects:1. In the traditional Gaussian mixture model (GMM), gaussian mixture number excessive will cause over smooth problems of synthesized speech. This paper proposes weighted sum of part of the conversion component functions instead of whole conversion functions in traditional GMM to improve the over smooth phenomenon. In order to avoid over smooth problem further, the warping function was designed to converse the LSF parameters directly, then the converted spectrum is adjusted by using the results of improved GMM method. The method can make voice spectral envelope conversion more accurate.2. The paper improves the traditional pitch detection method by introducing a new algorithm called RWAF. It is combined with the weighted candidate values to determine the final pitch. This method can improve pitch detection accuracy effectively.3. The traditional pitch transformation is only for short-term pitch with fixed mean conversion. To solve this problem, the paper provides a method to extract element segment characteristic parameters vector from pitch trajectory. An independent modeling based on GMM is used to train the extracted parameters, the obtained real-time dynamic conversion rules have more advantages than the fixed conversion rules to make the pitch transformation effect better.4. The voice conversion system combined with the above mothed is implemented. The converted voice is evaluated by both objective and subjective testing. Experimental results show that the system has better performance. It makes the speaker's individuality obvious, and also improves the converted voice quality effectively.