[New Model]: Vevo2 — Unified Controllable AR + Flow-Matching for Speech and Singing Voice Generation

The model to consider.

Vevo2 by Xueyao Zhang et al. (CUHK-Shenzhen / Amphion) is a unified, controllable framework for both speech and singing voice generation. The paper has been accepted to IEEE TASLP.

Asset	Location
Code	https://github.com/open-mmlab/Amphion/tree/main/models/svc/vevo2 (Amphion)
Checkpoint	https://huggingface.co/RMSnow/Vevo2
Paper	https://arxiv.org/abs/2508.16332
Demos	https://versasinger.github.io/

The closest model vllm-omni already supports.

The AR + flow-matching pattern in Vevo2 mirrors what we already run for codec-token TTS:

• Stage 1 (AR content-style modelling) corresponds to the Talker-style decoder-only LM in Qwen3-TTS / Qwen3-Omni.
• Stage 2 (flow-matching acoustic) corresponds to the diffusion / flow decoder used by current Code2Wav-adjacent stages.

What's your difficulty of supporting the model you want?

Architecture overview (from paper abstract + project page)

• Two specialized tokenizers:
  • Unified prosody tokenizer: captures prosody / melody from speech, singing, and instrumental audio without music notation.
  • Unified content-style tokenizer: encodes linguistic content, prosody, and style for both speech and singing while enabling timbre disentanglement.
• Two-stage generation:
  • AR content-style stage: controllability over text, prosody, and style; uses both explicit and implicit prosody learning to bridge speech and singing.
  • Flow-matching acoustic stage: independent timbre control.
• Multi-objective post-training: joint intelligibility + prosody-similarity alignment.

Integration notes

• The AR stage is the natural target for vllm-omni's existing AR runner pattern; the flow-matching stage would slot into a downstream stage similar to how diffusion vocoders are wrapped today.
• Code lives inside the Amphion repo, so packaging will involve isolating the Vevo2 module + its tokenizers from the larger Amphion codebase, similar to how other Amphion-derived work is already integrated upstream.
• The single framework covers both speech and singing — a model registry entry should make this dual capability explicit.

Open questions before integration

• Confirm checkpoint license on RMSnow/Vevo2 and Amphion's per-model license terms.
• Confirm the prosody tokenizer's input/output rate and any music-notation-free conditioning options for the SVS path.
• Determine whether the AR stage emits codec tokens compatible with an existing decoder, or requires a Vevo2-specific decoder stage.

Use case and motivation

1. Unified speech + singing: a single model covers both, which is a natural fit for vllm-omni's omni positioning.
2. AR-first stage is the most vllm-omni-native singing model in the current open-source landscape.
3. TASLP-accepted research with a public Amphion-hosted reference implementation and a public checkpoint.
4. Extends vllm-omni beyond speech: gives users a research-grade SVS option to complement product-grade SVS work (e.g. SoulX-Singer).

References

• Paper: https://arxiv.org/abs/2508.16332
• Code (Amphion): https://github.com/open-mmlab/Amphion/tree/main/models/svc/vevo2
• Checkpoint: https://huggingface.co/RMSnow/Vevo2
• Demo / project page: https://versasinger.github.io/

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.