MeanVC: Lightweight and Streaming Zero-Shot Voice Conversion via Mean Flows

MeanVC is a lightweight and streaming zero-shot voice conversion system that enables real-time timbre transfer from any source speaker to any target speaker while preserving linguistic content. The system introduces a diffusion transformer with chunk-wise autoregressive denoising strategy and mean flows for efficient single-step inference.

✨ Key Features

• 🚀 Streaming Inference: Real-time voice conversion with chunk-wise processing.
• ⚡ Single-Step Generation: Direct mapping from start to endpoint via mean flows for fast generation.
• 🎯 Zero-Shot Capability: Convert to any unseen target speaker without re-training.
• 💾 Lightweight: Significantly fewer parameters than existing methods.
• 🔊 High Fidelity: Superior speech quality and speaker similarity.

💾 Model Download

Use the following Python script to download the models into a local directory (e.g., ./checkpoints):

from huggingface_hub import snapshot_download

# Download all necessary models and components for MeanVC
snapshot_download(
    "ASLP-lab/MeanVC",
    allow_patterns=[
        "model_200ms.safetensors", # The trained MeanVC model weights
        "meanvc_200ms.pt",       # JIT-compiled model for real-time inference
        "fastu2++.pt",           # JIT-compiled ASR model
        "vocos.pt"               # JIT-compiled Vocos vocoder
    ],
    local_dir="./checkpoints", # Specify your target directory
    local_dir_use_symlinks=False
)

📜 License & Disclaimer

MeanVC is released under the Apache License 2.0. This open-source license allows you to freely use, modify, and distribute the model, as long as you include the appropriate copyright notice and disclaimer.

MeanVC is designed for research and legitimate applications in voice conversion technology. Users must obtain proper consent from individuals whose voices are being converted or used as references. We strongly discourage any malicious use including impersonation, fraud, or creating misleading audio content. Users are solely responsible for ensuring their use cases comply with ethical standards and legal requirements.

📄 Citation

If you find our work helpful, please cite our paper:

@article{ma2025meanvc,
  title={MeanVC: Lightweight and Streaming Zero-Shot Voice Conversion via Mean Flows},
  author={Ma, Guobin and Yao, Jixun and Ning, Ziqian and Jiang, Yuepeng and Xiong, Lingxin and Xie, Lei and Zhu, Pengcheng},
  journal={arXiv preprint arXiv:2510.08392},
  year={2025}
}

📧 Contact

If you are interested in leaving a message to our research team, feel free to email guobin.ma@mail.nwpu.edu.cn