LCM-SVC Demo

Abstract

Any-to-any Singing Voice Conversion (SVC) aims to transfer a target singer’s timbre to other songs using a short voice sample. Many SVC methods have achieved impressive results using diffusion models, which, however often suffer from high latency caused by inference steps. In this paper, we propose LCM-SVC, a latent consistency distillation (LCD) based latent diffusion model (LDM) to improve the inference speed. By leveraging significant improvements in timbre decoupling and sound quality offered by LDM, we can distill a pre-trained LDM, enabling one-step or multi-step inference while maintaining high performance. Results show that our method can significantly reduce the inference time and largely preserve the sound quality and timbre similarity.

Seen-to-Seen Singing Voice Conversion

Male-to-Male

Male-to-Female

Female-to-Male

Female-to-Female

Unseen-to-Unseen Singing Voice Conversion

Male-to-Male

Male-to-Female

Female-to-Male

Female-to-Female