CA-VC: A Novel Zero-Shot Voice Conversion Method With Channel Attention

In recent years, more and more zero-shot voice conversion algorithms have been proposed. However, tasks about adaptation and disentanglement of speaker and content information in speech are still challenging. In this paper, we propose a novel zero-shot voice conversion method with channel attention (CA) under the framework of variational auto-encoder. In detail, the model consists of content encoder, speaker encoder and decoder. CA plays two different roles in our method. In content encoder, CA with channel width constraint forms a learnable bottleneck to reduce speaker information and retain content information simultaneously. In decoder, CA is used to combine speaker information and content information. Objective and sub-jective evaluations show that the proposed method can perform voice conversion well and generate high quality converted speech.