Frequency warping for speaker adaption of text-to-speech synthesis

Abstract

Vocal tract length normalization (VTLN) is generally used in speech recognition for removing individual speaker characteristics. In this paper, we employ VTLN to speaker adaptation of speech synthesis. We propose a new frequency warping approach to reduce the spectrum distance between source and target speakers. The frequency warping function is based on a bilinear function and the warping factor is dynamically generated frame-by-frame. The warped spectra of source speaker are then converted to LSPs to train hidden Markov models (HMM). HMMs are further adapted by maximum likelihood linear regression (MLLR) with target warping approach can make the warped spectra of source speaker closer to target speaker and the resultant adapted HMMs have a better performance than the HMMs trained with unwarped spectra in term of voice naturalness and speaker similarity.