Integrating global variance of log power spectrum derived from LSPs into MGE training for HMM-based parametric speech synthesis

This paper presents a method to improve hidden Markov model (HMM) based parametric speech synthesis by integrating global variance (GV) of log power spectrum (LPS) derived from line spectral pairs (LSPs) into minimum generation error (MGE) model training. In order to alleviate the over-smoothing effect of the generated spectral structures, an LPS-GV based parameter generation method has been proposed. This method improved the naturalness of synthetic speech when LSPs were used as spectral features. However, it increased the complexity of parameter generation at synthesis time significantly. In this paper, we propose a method to integrate the distortions of LPS-GV derived from LSPs into the criterion of MGE model training in order to utilize LPSGV information at training time instead of at synthesis time. The experimental results show that this proposed method can achieve better naturalness of synthetic speech than the conventional MGE model training without loss of efficiency at synthesis time when LSPs are used as spectral features.