OmniVoice — single-file ComfyUI checkpoints

Single-file checkpoints of k2-fsa/OmniVoice (Qwen3 LLM + HiggsAudioV2 codec), repacked for ComfyUI-OmniVoice-Native. Config, codec config, and BPE tokenizer are embedded as safetensors metadata — no sidecar files needed.

Files

• omnivoice.safetensors (3.3GB, full quality) — most robust, recommended for tags + voice-design prompts.
• omnivoice-int8.safetensors (1.1GB, INT8-ConvRot) — same quality on plain prompts, smaller download. ~86% top-1 agreement with full.

Use

1. Install the pack: ComfyUI Manager → search OmniVoice, or git clone https://github.com/Aero-Ex/ComfyUI-OmniVoice into ComfyUI/custom_nodes/. No pip install needed.
2. Put a file from this repo into ComfyUI/models/omnivoice/.
3. Load OmniVoice Model → OmniVoice Generate. For a designed voice, add OmniVoice Voice Design → instruct.

Defaults: 32 steps, guidance 2.0, 25 tok/s. num_steps 16 roughly halves generation time. Texts over ~30s are auto-split into ~15s chunks and cross-faded.

Notes

• Voice clone: ref_audio + ref_text (transcript required, no ASR). One voice per Generate call — multi-speaker dialogue = one Generate per speaker, then concat.

• Voice design: gender, age, pitch, whisper, English accent / Chinese dialect via the instruct string.

• Vocalization tags (exact spelling, spoken by the model): [laughter], [sigh], [question-en], [question-ah], [question-oh], [question-ei], [question-yi], [confirmation-en], [surprise-ah], [surprise-oh], [surprise-wa], [surprise-yo], [dissatisfaction-hnn].

  Wait, what [surprise-ah]? You finished it already [question-oh]? Ha [laughter]! I knew you could do it [confirmation-en]. Phew [sigh], now we can finally rest.
  

• Keep normal punctuation spacing around tags (! Ha, . Phew, never !Ha). Crushed punctuation derails takes, especially on INT8.

• Same seed can give different takes across checkpoints — test seeds 1–3 per setup and pin the keeper.

• No output post-processing — use standard ComfyUI audio nodes for gain/fades/trim.

Credit / license

Model: k2-fsa OmniVoice. This repo only repacks the weights into single-file ComfyUI checkpoints; see the upstream repo and paper for training details.