YingMusic-SVC: Real-World Robust Zero-Shot Singing Voice Conversion with Flow-GRPO and Singing-Specific Inductive Biases

github:YingMusic-SVC

The difficulty grading benchmark for SVC. Each sample provides a clean vocalist (lead)/harmony (back)/ full song (mix)/ full vocal (mix_vocal) and the lead vocalist obtained using our self-developed separation model (ourlead).

The metadata records the gender of the singer for each sample, as well as the presence of echo and reverberation in the lead vocals

As a reference, our test set used the following Settings for SVC:

1. GT Leading: lead
2. Mix Vocal: mix_vocal
3. Ours Leading: ourlead

Short Intro

Singing voice conversion (SVC) aims to render the target singer’s timbre while preserving melody and lyrics. However, existing zero-shot SVC systems remain fragile in real songs due to harmony interference, F0 errors, and the lack of inductive biases for singing. We propose YingMusic-SVC, a robust zero-shot framework that unifies continuous pre-training, robust supervised fine-tuning, and Flow-GRPO reinforcement learning. Our model introduces a singing-trained RVC timbre shifter for timbre–content disentanglement, an F0-aware timbre adaptor for dynamic vocal expression, and an energy-balanced rectified flow matching loss to enhance high-frequency fidelity. Experiments on a graded multi-track benchmark show that YingMusic-SVC achieves consistent improvements over strong open-source baselines in timbre similarity, intelligibility, and perceptual naturalness—especially under accompanied and harmony-contaminated conditions—demonstrating its effectiveness for real-world SVC deployment.