OmniVoice

ComfyUI-OmniVoice

A native, dependency-free port of the OmniVoice TTS model (Qwen3 LLM + HiggsAudioV2 codec) for ComfyUI. No extra pip packages — runs on stock ComfyUI (INT8 checkpoints use ComfyUI's built-in quantization stack).

Nodes

• Load OmniVoice Model — loads a single-file checkpoint (models/omnivoice/*.safetensors). Auto-detects float and INT8-ConvRot checkpoints.
• OmniVoice Voice Design — per-category dropdowns (gender, age, pitch, style, accent, dialect) that build the voice-design string. Connect to Generate's instruct.
• OmniVoice Generate — text-to-speech. Optional voice clone (ref_audio + ref_text), auto duration estimation, chunked long-form generation (texts over ~30s are split into ~15s chunks and cross-faded, near-constant VRAM).

Install

Only two ways:

1. ComfyUI Manager — Manager → Custom Nodes Manager → search OmniVoice → Install. Restart ComfyUI.
2. Manual — clone this repo into ComfyUI/custom_nodes/:

   cd ComfyUI/custom_nodes
   git clone https://github.com/Aero-Ex/ComfyUI-OmniVoice
   

   Restart ComfyUI. No installtion needed.

Checkpoints

Download a ready single file from Aero-Ex/OmniVoice into models/omnivoice/:

• omnivoice.safetensors (3.3GB, full quality)
• omnivoice-int8.safetensors (1.1GB, INT8-ConvRot, ~86% top-1 agreement with full)

Vocalization tags

Type these inline in the prompt, exact spelling in brackets — the model speaks them:

• [laughter], [sigh]
• [question-en], [question-ah], [question-oh], [question-ei], [question-yi], [confirmation-en]
• [surprise-ah], [surprise-oh], [surprise-wa], [surprise-yo]
• [dissatisfaction-hnn]

Example:

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 or .Phew). Crushed punctuation creates rare token pieces that derail takes, especially on the INT8 checkpoint (verified: same prompt babbles 5/5 unspaced, speaks spaced). Anything else in brackets is read out as literal text.

Usage

Load OmniVoice Model → OmniVoice Generate. For a designed voice, add OmniVoice Voice Design → instruct.

Notes:

• ref_text is required whenever ref_audio is connected (no ASR auto-transcription).
• No output post-processing is applied — use standard ComfyUI audio nodes for gain/fades/trim.
• Checkpoint picks: bf16 for tags + voice-design prompts (most robust); int8 for plain prompts (same quality, smaller). Same seed can give different takes across checkpoints — pin a keeper seed per setup.