Composite Wavelet Model for Stability-Oriented Speech Synthesis from Cepstral Features

This paper discusses a stability-oriented vocoder based on Gabor wavelet approximation of the source signal for statistical speech synthesis. In conventional vocoders with recursive filters, the filter gain characteristics often cause degradations in the sound quality due to unstable behavior of recursive filters affected by sharp resonances driven by a particular overtone in the excitation signal. To cope with this problem, we have proposed Composite Wavelet Model (CWM) to avoid filter-caused problems and have made several improvements as a vocoder. Based on non-recursive filters, it enables synthesizing stable speech which is robust to changes in F-0 parameter. In this paper, we further discuss the optimal number of mixture components to improve the synthetic speech quality to determine them through subjective experimental evaluations and report them on the result of incorporating in an HMM-based speech synthesis system. Objective experimental evaluations confirmed the improved stability in the amplitude of the synthetic speech.