Voice conversion of different ages using universal background model groups of short-time spectra and prosodic features

For the voice conversion of different ages, a method using Universal Background Model Groups(UBMG) of short-time spectra and prosodic features is proposed. In spectrum aspect, Gaussian Mixture Model(GMM) is trained for every speaker after extracting linear predictive cepstrum coefficients, then the speakers in the same age period are clustered based on their voice similarity, and each cluster is further trained to be a UBM of spectrum distribution.Finally, an UBM group and corresponding spectrum conversion functions are obtained in each age period. Formants adjustment is further used after spectrum conversion. Furthermore, fundamental frequency and speech rate are modeled by single Gaussian and average duration rate respectively to derive their conversion functions in the aspect of prosodic features. The results of objective and subjective evaluation experiments such as ABX and MOS show that the proposed method has a distinct advantage compared with conventional bilinear method and its change rate of log-likelihood ratio increases by 4% compared with single UBM method. The results show the proposed method can make the converted speech more close to the speech of target age period with good speech quality while the performance has been improved evidently compared with conventional methods.