Probabilistic Approach for Speaker Transformation

A probabilistic approach of speaker transformation is proposed in this paper to make the speech of a source speaker sound like uttered by a target speaker. Speaker individuality transformation is achieved by altering characteristics of the speech spectrum and the supersegmental information such as fundamental pitch frequency. The main advantage of this scheme lies in the aspect of not only having considered the statistical property of both the source and target speech spectrum but also the relationship between them under a cross correlational model. And to make sure that the transformed speech signals are perceptually closer to the target speaker, prosody modification is also involved. The proposed scheme is evaluated using both subjective and objective measures. The experimental results show that the transformation system put forward is capable of effectively transforming speaker identity whilst the converted speech maintains high quality. And the whole performance is evaluated to be superior to the conventional vector quantization (VQ) based method.