Voice morphing based on spectral features and prosodic modification

Abstract

This paper is aimed at morphing the speech uttered by a source speaker in a manner that it seems to be spoken by another target speaker - a new identity is given while preserving the original content. The proposed method transforms the vocal tract parameters and glottal excitation of the source speaker into target speaker's acoustic characteristics. It relates to the development of appropriate vocal tract models that can capture information specific to the speaker and estimate the model parameters that closely relate to the model of the target speaker. It detects the pitch, separates the glottal excitation and vocal tract spectral features. The glottal excitation of the source is taken, voice/un-voice decision is made, the prosody information is found, PSOLA (Pitch Synchronous OverLap Add) is used to modify the pitch, the spectral features are found, and finally speech is modified using target spectral features and prosody. The subjective experiment shows that the proposed method improves the quality of conversion and contains the original vocal and glottal characteristics of the target speaker.