Speech Analysis/Synthesis by Gaussian Mixture Approximation of the Speech Spectrum for Voice Conversion

Voice conversion typically employs spectral features to convert a source voice to a target voice. In this paper, we propose a simple method of fitting the STRAIGHT spectrum with Gaussian mixture (GM) models for speech analysis/synthesis and spectral modification. The mean values of the Gaussians are pre-determined based on Mel-frequency spacing. The standard deviations are also adaptively adjusted using the constant-Q principle and the spectrum amplitudes. Finally, the weights of the Gaussians are determined by sampling the log-spectrum at Mel-frequencies. The proposed analysis/synthesis method (MFLS-GM) is employed for speech analysis/synthesis and voice conversion. Subjective evaluations employing MOS and ABX demonstrate superior performance of the voice conversion using the MFLS-GM compared to systems employing MFCC features. The computation cost of the proposed analysis/synthesis method is also much lower than those based on MFCC.