Skip to content

Add native support for OpenBMB VoxCPM2 #47695

Description

@sushmanthreddy

Model description

VoxCPM2 is a 2B-parameter, tokenizer-free diffusion-autoregressive text-to-speech model developed by OpenBMB. Instead of relying on discrete speech tokens, it generates continuous speech representations through a
hierarchical architecture built around LocEnc, TSLM, RALM, and LocDiT components, with a MiniCPM-4 backbone and AudioVAE V2.

Key features:

  • Multilingual Speech Generation: Supports 30 languages and 9 Chinese dialects without requiring explicit language tags.
  • Natural-Language Voice Design: Creates new voices from text descriptions of attributes such as age, tone, emotion, gender, and speaking pace.
  • Controllable Voice Cloning: Clones a voice from a short reference recording while allowing style and emotion instructions.
  • High-Fidelity Continuation Cloning: Uses reference audio and its transcript to preserve timbre, rhythm, emotion, and speaking style.
  • Tokenizer-Free Architecture: Generates continuous speech representations without an external discrete speech tokenizer.
  • High-Quality Audio: Accepts 16kHz reference audio and produces 48kHz output through the asymmetric AudioVAE V2.
  • Streaming Generation: Supports real-time streaming text-to-speech generation.
  • Open Implementation: The implementation and weights are available under the Apache-2.0 license.

The current official loading path uses the external voxcpm package.

VoxCPM2 has strong public adoption. As of August 1, 2026, the Hugging Face checkpoint reports 1,005,747 downloads in the previous month and is used by 100 public Hugging Face Spaces. Its model tree also includes 7 adapters, 23 fine-tunes, and 9 quantizations, showing active community use and interest.

Would the maintainers be open to adding native VoxCPM2 support to Transformers?

If this model is within scope and nobody is already working on it, I would be happy to contribute the implementation and track the pull request through this issue. The exact native scope—configuration, model components, audio processing, checkpoint conversion, and generation API—can be agreed here before implementation.

Open source status

  • The model implementation is available
  • The model weights are available

Provide useful links for the implementation

[1] VoxCPM2 technical report: https://arxiv.org/abs/2606.06928
[2] Hugging Face checkpoint: https://huggingface.co/openbmb/VoxCPM2
[3] Original implementation: https://github.com/OpenBMB/VoxCPM
[4] Official documentation: https://voxcpm.readthedocs.io
[5] Source code: https://github.com/OpenBMB/VoxCPM/tree/main/src/voxcpm
[6] Apache-2.0 license: https://github.com/OpenBMB/VoxCPM/blob/main/LICENSE

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions