Voice Conversion using structrued Gaussian Mixture Model

Gaussian Mixture Model (GMM) is commonly used in voice conversion. However, traditional GMM based voice conversion usually extracts a conversion function from parallel corpus, which greatly limits the application of the technology. In an attempt to overcome this drawback, structured Gaussian Mixture Model (SGMM) is applied to model the speaker´s acoustic feature distribution. In particular, two speakers´ isolated SGMMs are aligned based on Acoustic Universal Structure (AUS) theory. Then the conversion function is extracted from two aligned SGMMs in a manner similar to conventional method. The subjective listening tests indicate that the proposed method achieves equivalent speech quality and speaker individuality compared with conventional method.