A speech parameter generation algorithm using local variance for HMM-based speech synthesis

This paper proposes a parameter generation algorithm using local variance (LV) constraint of spectral parameter trajectory for HMM-based speech synthesis. In the parameter generation process, we take account of both the HMM likelihood of speech feature vectors and a likelihood for LVs. To model LV precisely, we use dynamic features of LV with context-dependent HMMs. The objective experimental results show that the proposed technique can generate a better spectral trajectory in terms of the spectral and LV distortions than a conventional technique with global variance (GV) constraint. The subjective experimental results also show that the proposed technique significantly improve the reproducibility of the synthetic speech than the conventional one.