(A) nonlinear voice conversion method using gaussian mixture model

Abstract

Voice conversion is the technique for modifying the speech signal of a source speaker so that it sounds as if it had been uttered by another target speaker. There have been many methods being proposed for voice conversion, among which the linear transformation methods using the Gaussian Mixture Models (GMM) have been shown to outperform the others. In this thesis, we identify the problem of the GMM-based linear transformation methods, the over-smoothing effect of the converted speech, and propose a new GMM-based nonlinear transformation method using Radial Basis Function networks. Our system is implemented in the context of the Harmonic plus Noise Model to achieve high quality modification of speech. Our experiments show that our system succeeds in converting speech and our nonlinear method outperforms linear transformation methods for large number of mixture components.