Speech parametric representation for hmm based speech synthesis

Abstract EN

MM based speech synthesis, where speech parameters are generated directly from HMM models, is a relatively new speech synthesis technique.

In this paper, we propose a parametric representation for the HMM based speech synthesis.

In this representation, a multi-band excitation model is used, and a number of samples extracted from the spectral envelop are used as spectral parameters.

In synthesis, voiced and unvoiced speech parts are mixed according to bands voicing parameters.

The voiced part is generated based on a harmonic sinusoidal model, and unvoiced part is generated as white noise filtered using short time fourier transform.

Experimental results show that the proposed parametric representation improved the quality.