Aero-Ex / ComfyUI-Vibevoice

VibeVoice 7B for ComfyUI

Single-file checkpoints of Microsoft VibeVoice 7B (VibeVoice-Large) for the native ComfyUI node pack Aero-Ex-ComfyUI-Vibevoice . They do not load in the original vibevoice package or in transformers.

Put one in ComfyUI/models/text_encoders/ and load it with Load VibeVoice Model .

Conversion

The weights are the original VibeVoice-Large weights, repacked into one file with the tokenizer and config:

• attention Q/K/V and MLP gate/up projections merged,
• LM head pruned to the four speech control tokens,
• INT8 file quantized with comfy-kitchen's TensorWiseINT8 ConvRot layout from the bf16 file.

Checked against the transformers VibeVoice implementation: identical prompt and generated speech tokens with matched noise.

Usage

Speaker 1: Welcome back to the podcast.
Speaker 2: Thanks, it is great to be here.

Connect one 5–15 s voice sample per speaker. The model was trained on up to 4 speakers.

License and responsible use

MIT, as the original release. VibeVoice can clone voices: only use voices you have permission to use, and disclose AI-generated audio. See Microsoft's model card for the original risks and limitations.