Research On Technologies Of Voice Conversion Based On Gaussian Mixture Model

Voice conversion is a technique that modifies a source speaker?s speech to be perceived as if a target spearker?s, so that the source speaker?s voice sounds like the target speaker?s voice. It is an exciting new branch of speech signal processing. With the improvement of the people's quality of life, we are not only request the understanding of our voice, but also emphasize the individual characters of speech. So, the research of this technology both has great significance in theory and the value of application.The main work of this paper:First, discuss the model and system of vocal sound,the basic theory and method of voice conversion techniques, and the speech parameters. Study on the Gaussian mixture model (GMM) for performing spectral conversion between speakers .Second, the converted spectral are excessively smoothed by statistical Gaussian mixture model (GMM) algorithm. In order to address this problem, we propose a conversion method (GMM-GV) considering global variance feature. The emulation demonstrates that GMM-GV algorithm can overcome the problem effectively.Third, the spectral discontinuities due to independent mapping of subsequent frames. In order to address those problems, an enhanced conversion method(GMM-GV-Viterbi) incorporating frame selection algorithm is presented.Forth, F0 contour reflect individuality of speaker also. We propose an improved method (GMM-Viterbi) to modify prosody. The experiment indicate that the performance of proposed method is better than the statistical method(GMM).