Voice conversion using support vector regression

A new voice conversion method based on support vector regression (SVR) is proposed, and the mapping abilities of a multi-dimensional SVR are exploited to perform the mapping of spectral features of a source speaker to that of a target speaker. A novel mixed kernel is presented to improve the mapping performance, the correlations between frames of the source speaker are considered to overcome the discontinuities of converted speech, and an adaptive median filter is adopted in the conversion phase to smooth the converted spectral parameter trajectory. Experimental results show that the proposed method outperforms the state-of-the-art Gaussian mixture model based method, it can achieve high similarity between converted and target speakers, and has good quality and naturalness.