MultipleContentsSVC Demo Page

Performance of Our Proposed SVC System

The compared methods are as follows:

FastSVC: It is an official baseline provided by the SVC Challenge 2023 (SVCC 2023). It is a GAN-based wav-to-wav model, which utilizes a pretrained conformer to extract content features (similar to WeNet).
SoVITS: It is developed by a popular SVC open-source community. It adopts a flow-based acoustic decoder and utilizes ContentVec as the content features.
Ours (W+C): It adopts a diffusion-based acoustic decoder and uses both Whisper and ContentVec as the content features.
Ours (W+W+C):It adopts a diffusion-based acoustic decoder and uses the WeNet, Whisper, and ContentVec as the content features.

We adopt three tasks to evaluate the conversion performance: (1) Conversion from M4Singer to Opencpop (in Chinese; in-domain conversion), (2) SVCC in-domain conversion task (in English), and (3)SVCC cross-domain conversion tasks (in English).