A Bayesian approach to HMM-based speech synthesis

This paper proposes a new framework of speech synthesis based onthe Bayesian approach. The Bayesian method is a statistical techniquefor estimating reliable predictive distributions by marginalizingmodel parameters. In the proposed framework, all processes forconstructing the system can be derived from one single predictivedistribution which represents the basic problem of speech synthesisdirectly. Using HMM as the likelihood function and assuming someapproximations, it can be regarded as an application of the variationalBayesian method to the HMM-based speech synthesis. Experimentalresults show that the proposed method outperforms theconventional one in a subjective test