MoCoVC: Non-parallel Voice Conversion with Momentum Contrastive Representation Learning

Non-parallel voice conversion with deep neural networks often disentangle speaker individuality and speech content. However, these methods rely on external models, text data, or implicit constraints for ways to disentangle. They may require learning other models or annotating text, or may not understand how latent representations are acquired. Therefore, we propose voice conversion with momentum contrastive representation learning (MoCoVC), a method of explicitly adding constraints to intermediate features using contrastive representation learning, which is a self-supervised learning method. Using contrastive representation learning with transformations that preserve utterance content allows us to explicitly constrain the intermediate features to preserve utterance content. We present transformations used for contrastive representation learning that could be used for voice conversion and verify the effectiveness of each in an experiment. Moreover, MoCoVC demonstrates a high or comparable performance to the vector quantization constrained method in terms of both naturalness and speaker individuality in subjective evaluation experiments.