Speaker Transformation Based On Prosodic With Short Time Spectrum

Speaker transformation means that developing a voice conversion system that convert the speech signal uttered by a source speaker to a speech signal which have the voice characteristics of the target speaker.But the semanteme should not be changed.This research is relative to so many domains such as speech analysis, speech coding, speech synthesis, speech enhancement, speech recognition.It also has important application in education, military affairs and so on.The research is as follows:(1) Bring forward a new method of speaker transformation using prosodic features with short time spectrum. By using a GMM conversion function, the LSF vectors and pitch period of every frame belong to original signal are joined into an eigenvector, and then it is conversed to be as features of target spekaer., The LPC residuel of source speaker is also modified to that of target speker by a pitch conversion model.(2) Evaluated the performance of conversion model with different parameters. Subjective test ABX , MOS and spectrum distortion measure as objective test are used. The proposed method is also compared with the traditional method in 4 aspects. Subejective and obejective test show that the improved method has the better conversion performance compared to the traditional methods.The transformed voices sound more naturely, the intelligibility and orientation have also been improved.