S2CD: Self-heuristic Speaker Content Disentanglement for Any-to-Any Voice Conversion

In this paper, we propose a Selfheuristic Speaker Content Disentanglement (S2CD) model for any to any voice conversion without using any external resources, e.g., speaker labels or vectors, linguistic models, and transcriptions. S2CD is built on the disentanglement sequential variational autoencoder (DSVAE), but improves DSVAE at the model architecture level from three perspectives. Specifically, we develop different structures for speaker and content encoders based on their underlying static/dynamic property. We further propose a generative graph, modelled by S2CD, so as to make S2CD well mimic the multi-speaker speech generation process. Finally, we propose a self-heuristic way to introduce bias to the prior modelling. Extensive empirical evaluations show the effectiveness of S2CD for any to any voice conversion.