Warped source spectrum for voice conversion and similarity

Abstract

A hybrid voice conversion method was developed that combines frequency warping and unit selection. The system first generates the warped source spectrum by frequency warping and then uses the warped source spectrum as a target to select the target speaker's real spectrum. The target speaker's spectrum then substituted for part of the warped source spectrum before reconstructing the converted speech to improve the similarity over the entire spectrum. Evaluations show that the hybrid voice conversion method improves the similarity more than frequency warping alone.