YingMusic-SVC: Real-World Robust Zero-Shot Singing Voice Conversion with Flow-GRPO and Singing-Specific Inductive Biases
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.
-
Three‑Stage Training Pipeline
- CPT: Continuous Pre-Training with singing‑trained modules
- SFT: Robust Supervised Fine-Tuning with F0 perturbation & harmony augmentation
- RL (Flow‑GRPO): Multi-reward reinforcement learning for perceptual quality
-
Singing-Specific Inductive Biases
- 🎼 RVC-based Timbre Shifter (trained on 120 singers)
- 🎚️ F0‑Aware Fine-Grained Timbre Adaptor
- 🔊 Energy-balanced Flow Matching Loss (enhanced high-frequency details)
- 2025-11-26: Released our accompany separator inference CLI and model ckpt
- 2025-11-26: Released gradio app for easy try
- 2025-11-25: Released technical report
- 2025-11-25: Initial YingMusic-SVC inference CLI
- 2025-11-25: Released model checkpoint
- 2025-11-25: Released multi-track benchmark
git clone https://github.com/GiantAILab/YingMusic-SVC.git
cd YingMusic-SVC
conda create -n ymsvc python=3.10
conda activate ymsvc
pip install -r requirements.txt
# install ffmpeg & sox
sudo apt update
sudo apt install -y sox libsox-fmt-all
sudo apt install -y ffmpegcd accom_separation
bash infer.sh
bash my_infer.shpython gradio_app.pyWe provide a graded difficulty benchmark, derived from 100+ multi-track studio songs:
| Level | Description |
|---|---|
| GT Leading | Clean studio lead vocals |
| Mix Vocal | Lead + harmony contamination |
| Ours Leading | Extracted via our Band RoFormer separator |
| Model | Description | Link |
|---|---|---|
| YingMusic-SVC-full | RL-enhanced final model |  |
| our BR separator | Our accompany separation model |  |
- our stem-separator inference CLI & model ckpt
- develop gradio app for YingMusic-SVC
- benchmark one-click eval script
This project is built upon:
If you use YingMusic‑SVC for research, please cite:
@article{chen2025yingmusicsvc,
title={YingMusic-SVC: Real-World Robust Zero-Shot Singing Voice Conversion with Flow-GRPO and Singing-Specific Inductive Biases},
author={Chen, Gongyu and Zhang, Xiaoyu and Weng, Zhenqiang and Zheng, Junjie and Shen, Da and Ding, Chaofan and Zhang, Wei-Qiang and Chen, Zihao},
journal={arXiv preprint arXiv:2512.04793},
year={2025}
}
Our code is released under MIT License.