ComfyUI custom node for OmniVoice TTS and voice cloning.

Upstream project:

• https://github.com/k2-fsa/OmniVoice

Download Model Here:

• https://huggingface.co/k2-fsa/OmniVoice

This node is built with only the necessary model forward core process in mind, so the nodes in this repo are already the final ones. I don’t like cluttering ComfyUI with unnecessary node mappings, so updates will only include bug fixes or truly urgent and necessary new nodes. If needed, I will create them.

Check what your ComfyUI environment is using:

pip list | grep transformer

Why: some models and libraries might still heavily depend on 4.5X HF Transformers.

Direct manual clone:

1. git clone https://github.com/komikndr/omnivoice_comfy inside ComfyUI/custom_nodes
2. cd omnivoice_comfy
3. pip install -r requirements.txt

ComfyUI manager:

1. comfy node install omnivoice_comfy

2. Put the OmniVoice weights in ComfyUI/models/tts/omnivoice/.

Expected layout:

ComfyUI/
  models/
    tts/
      omnivoice/
        model.safetensors
        audio_tokenizer.safetensors

You only need to place the two .safetensors files in the folder above. The node already includes the required tokenizer and config assets.

Loads:

• OmniVoice Model
• Audio Tokenizer Model

The loader builds a local runtime snapshot from the embedded config assets and the two selected weight files.

Inputs:

• text for the target speech
• optional instruct
• optional ref_audio and ref_text for voice cloning

If you use ref_audio, you must also provide ref_text.

• Whisper auto-transcription is disabled. Voice cloning requires ref_text.
• If you want voice cloning, install https://github.com/yuvraj108c/ComfyUI-Whisper or another similar workflow/pipeline that auto-transcribes the source audio. OmniVoice requires the transcript of the source audio. You can manually transcribe a 3 second clip, but that gets tedious in batch processing.
• The node uses files from ComfyUI/models/tts/omnivoice/ and builds a symlinked runtime snapshot.
• If symlink creation fails on your system, use a full HuggingFace-style OmniVoice folder instead.

• This repo is build with the help of Qwen 3.5 9B and embeddinggemma-300m to store the original code into vector store for fast retrieval (most of my time in coding wasted on code repo search)