Voice conversion with linear prediction residual estimaton

The work presented here shows a comparison between a voice conversion system based on converting only the vocal tract representation of the source speaker and an augmented system that adds an algorithm for estimating the target excitation signal. The estimation algorithm uses a stochastic model for relating the excitation signal to the vocal tract features. The two systems were subjected to objective and subjective tests for assessing the effectiveness of the perceived identity conversion and the overall quality of the synthesized speech. Male-to-male and female-to- female conversion cases were tested. The main objective of this work is to improve the recognizability of the converted speech while maintaining a high synthesis quality.