An Improvement of HSMM-Based Speech Synthesis by Duration-Dependent State Transition Probabilities

摘要

In this paper, we propose an improvement of hidden semi- Markov model (HSMM) based speech synthesis system by duration- dependent state transition probabilities. In traditional HMM algorithm, the probability of the duration of a state decreases exponentially with time, which does not provide an adequate representation of the temporal structure of speech. To overcome this limitation, HSMM, which models explicitly the state duration distribution, was proposed. However, there is still an inconsistency. Although HSMM has explicit state duration probability distributions, the state transition probabilities are duration-invariant. In this paper, we introduce duration-dependent state transition probabilities, which are able to characterize the timescale distortion at particular instant of an utterance more effectively, into HSMM based speech synthesis system. Correspondingly we improve forward-backward algorithm and re-derive parameter re-estimation formulae. Experimental results show that the proposed method improves the naturalness of the synthesized speech.