NoiseVC Voise Conversion Demos

Abstract

Zero-shot Voice conversion (VC) is a challenging task that transforms voice from a target audio to source without losing linguistic contents, while both source and target speakers are unseen during training. Previous approaches require a pre-trained model or linguistic data to do the zero-shot conversion. Meanwhile, VC models with vector quantization (VQ) or instance normalization (IN) are able to disentangle contents from audios achieve a successful conversion. However, disentanglement in these models highly relies on heavily constrained bottleneck layers, therefore, the audio quality is drastically sacrificed. In this paper, we propose NoiseVC, an approach that can disentangle contents based on VQ and Contrastive Predictive Coding (CPC) together. Additionally, we perform Noise Augmentation to further enhance disentanglement capability. We conduct several experiments and demonstrate that NoiseVC has a strong disentanglement ability without huge quality sacrifice.

Unseen speakers

Source and target speakers have not been seen during training. Audio samples are randomly picked from eval dataset.

Female to Male

Male to Female

Female to Female

Male to Male

Seen speakers

Source and target speakers have been seen during training. Audio samples are randomly picked from eval dataset.

Female to Male

Male to Female

Female to Female

Male to Male