CASIA Voice Conversion System for the Voice Conversion Challenge 2020

Abstract

This paper presents our CASIA (Chinese Academy of Sciences, Institute of Automation) voice conversion system for the Voice Conversation Challenge 2020 (VCC 2020). The CASIA voice conversion system can be separated into two modules: the conversion model and the vocoder. We first extract linguistic features from the source speech. Then, the conversion model takes these linguistic features as the inputs, aiming to predict the acoustic features of the target speaker. Finally, the vocoder utilizes these predicted features to generate the speech waveform of the target speaker. In our system, we utilize the CBHG conversion model and the LPCNet vocoder for speech generation. To better control the prosody of the converted speech, we utilize acoustic features of the source speech as additional inputs, including the pitch, voiced/unvoiced flag and band aperiodicity. Since the training data is limited in VCC 2020, we build our system by combining the initialization using a multi-speaker data and the adaptation using limited data of the target speaker. The results of VCC 2020 rank our CASIA system in the second place with an overall mean opinion score of 3.99 for speaker quality and 84% accuracy for speaker similarity.