A Bayesian Approach to Hidden Semi-Markov Model Based Speech Synthesis

This paper proposes a Bayesian approach to hidden semi-Markov model (HSMM) based speech synthesis. Recently, hiddenMarkov model (HMM) based speech synthesis based on theBayesian approach was proposed. The Bayesian approach is astatistical technique for estimating reliable predictive distributionsby treating model parameters as random variables. In theBayesian approach, all processes for constructing the system arederived from one single predictive distribution which exactlyrepresents the problem of speech synthesis. However, there isan inconsistency between training and synthesis: although thespeech is synthesized from HMMs with explicit state durationprobability distributions, HMMs are trained without them. Inthis paper, we introduce an HSMM, which is an HMM withexplicit state duration probability distributions, into the HMMbasedBayesian speech synthesis system. Experimental resultsshow that the use of HSMM improves the naturalness of thesynthesized speech