Voice conversion based on joint pitch and spectral transformation with component group-GMM

Spectral and pitch are two most important features in voice conversion which including a majority of speaker identity information. Some researchers use the GMM (Gaussian mixture model) to model the joint spectral and pitch. But these two features have the discrepancy of unit and meaning, so should do some processing before training the model. In this paper, a new framework CG-GMM (component-group GMM) is used for the joint pitch and spectral transformation. Experiments are setup and compared with the previous approach of voice conversion. The converted speeches indicate satisfactory speech quality and speaker identifiability. Meanwhile the speaking style is much like to the target speaker.