ComfyUI-OmniVoice_CRT

ComfyUI-OmniVoice

ComfyUI custom node wrapper for k2-fsa/OmniVoice, a multilingual zero-shot TTS model supporting voice cloning, voice design, and automatic voice generation.

Source

• OmniVoice repo: https://github.com/k2-fsa/OmniVoice
• OmniVoice model: https://huggingface.co/k2-fsa/OmniVoice

Features

• Voice cloning via reference audio/transcript prepared in the load node pipe
• Voice design (instruct prompt)
• Automatic mode selection based on reference audio presence
• Automatic first-run model download from Hugging Face into Comfy model cache

Installation

1. Install via ComfyUI Manager

   • Search for ComfyUI-OmniVoice_CRT and install.

   OR

   Clone manually into ComfyUI/custom_nodes and install dependencies:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/ComfyUI-OmniVoice_CRT.git
cd ComfyUI-OmniVoice_CRT
install_omnivoice_safe.bat

The BAT installer auto-detects python_embeded\\python.exe and installs:

• omnivoice>=0.1.0 with --no-deps (prevents Torch/Torchaudio override)
• huggingface_hub>=1.3.0,<2.0

2. Restart ComfyUI.

Usage Notes

• For voice cloning:
  • connect reference audio (optional) (and optional transcription override) to OmniVoice Load Model
• For voice design:
  • do not connect reference audio
  • set instruct (example: female, low pitch, british accent), or leave empty for fallback

Troubleshooting

• Model download fails
  • Check internet/Hugging Face access.
• ASR auto-transcription fails
  • Connect reference audio to OmniVoice Load Model and ensure network access for Whisper model download.

Not Implemented

• Batch multi-utterance generation in a single node execution (this wrapper generates one output audio at a time).