VQStarGAN: An Unseen-to-Unseen, Zero-Shot, Non-Parallel, Unsupervised Voice Conversion

Voice conversion (VC) is a technique that aims to transform the speech of one speaker into the speech of another while preserving the content. Existing solutions, such as StarGANv2-VC and VQ-based VC (VQVC), have focused on specific conversion scenarios like seen-to-seen and one-shot conversion, respectively. We propose the Vector Quantization Based StarGAN (VQStarGAN) model to conduct the unseen-to-unseen voice conversion. Our model adopts the basic framework of StarGAN in order to stabilize the training and produce natural-sounding speech. We have also incorporated the Vector Quantization (VQ) component to separate the content and speaker information in order to conduct unseen-to-unseen voice conversion. Subjective and objective evaluations of unseen-to-unseen voice conversion reveal that VQStarGAN achieves superior performance compared to StarGANv2-VC and VQVC.