Modulation Spectrum-based Postfiltering of Synthesized Speech in the Wavelet Domain

Abstract

This thesis presents a wavelet-domain measure used in postfiltering applications. Quality of HMM-based (hidden Markov model-based) parametric speech synthesis is degraded due to the over-smoothing effect, where the trajectory of generated speech parameters is smoothed out and lacks dynamics. The conventional method uses the modulation spectrum (MS) to quantify the effect of over-smoothing by measuring the spectral tilt in the MS. In order to enhance the performance, a modified version of the MS called the scaled modulation spectrum (SMS), which essentially separates the MS in different bands, is proposed and utilized in postfiltering. The performance of two types of wavelets, the discrete wavelet transform (DWT) and the dual-tree complex wavelet transform (DTCWT), are evaluated. We also extend the SMS into a hidden Markov tree (HMT) model, which represents the interdependencies of the coefficients. Experimental results show that the proposed method performs better.