An HMM Based Pitch-Contour Generation Method for Mandarin Speech Synthesis

Abstract

In this paper, a method is proposed to generate pitch-contours for Mandarin speech synthesis. In this method, an HMM (hidden Markov model) is used to model the pro- sodic states implicitly stayed and a syllable's pitch-contour is treated as an observation generated from a prosodic state. Such an HMM is called a syllable pitch-contour HMM (SPC-HMM). For training the SPC-HMM, we developed a feasible method to normalize a pitch-contour's height. After normalization, each training syllable's pitch-contour is vector quantized and represented with a VQ (vector quantization) code. Then, the VQ code and its adjacent syllables' lexical tones are combined to define an observation symbol for training the SPC-HMM. In the synthesis phase, a sentence-wide most prob- able observation symbol sequence is searched on the SPC-HMM using a dynamic pro- gramming algorithm proposed here. Then, the observation symbol found for a syllable is decoded to obtain its pitch-contour VQ code. We conducted testing experiments to de- termine the size of a pitch-contour codebook and the number of states for an SPC-HMM. The results indicate that setting the codebook size to eight and using six states are the best choices. Also, we conducted perception tests to compare the naturalness levels of synthetic speech files. The results show that the two generation modes for operating an SPC-HMM studied here are comparable to each other in naturalness level.