Expressive Voice Conversion Using Perturbation and Autoencoder Loss

Expressive voice conversion is to transform voice speech with emotional expression. Modeling emotional styles in expressive voice conversion is challenging. It is essential to eliminate the interference of factors such as the identity in emotional style modeling. This paper addresses this issue using an information perturbation method. During the feature extraction phase, perturbation functions are applied to factors such as formants and pitch in the speech signal to eliminate the interference of speaker information. This allows for separate modeling and extraction of content information and emotional information. We propose a general framework for expressive voice conversion, which aims to disentangle content information, speaker information, and emotional style information from the voice signal. Experimental results demonstrate the effectiveness of our proposed framework in both subjective and objective evaluations. It can extract highly expressive and understandable features from the source speech while maintaining similarity to the target speaker.