Large scale compression of HMM for statistical speech synthesis

A hidden Markov model (HMM) is used as the statistical model for the acoustic parameters in statistical speech synthesis. This paper presents a quantization method for large scale compression of the HMM based on the acoustic space distance in the speech models. The quality loss caused by the quantization is reduced by an optimal iteration procedure that optimizes the vector quantization codebook. Objective and subjective evaluations show that 90% of the scores indicate no significant reduction of the speech quality with a compression ratio of 0.06 using the proposed spectrum model compression method.