ChatterBox VC Loader (Manual)

ChatterBox VC Loader (Manual) is the lightweight side of the pack's ChatterBox story. Voice conversion - swapping one speaker's voice for another's - only needs one generative model, the S3Gen speech generator, so this loader has just two required inputs where the TTS loader has four. You supply s3gen.safetensors from ComfyUI/models/chatterbox/, pick a device, optionally load the built-in voice conditionals, and get a CHATTERBOX_VC model ready for the Manual VC node. It's the loader to use when you already have the weights and want deterministic, offline behavior.

How it works

The loader instantiates the S3Gen model, loads s3gen.safetensors into it, moves it to your chosen device, and wraps it in a small object exposing the model's sample rate. If you provide conds.pt, it loads the reference-conditioning dict for ChatterBox's built-in voices. That's the whole job - because voice conversion's second half (resynthesizing tokens under a target voice) is handled entirely inside the VC node itself using this one model. The wrapped object is returned as a CHATTERBOX_VC output.

The inputs

• s3gen - pick s3gen.safetensors from ComfyUI/models/chatterbox/
• device - auto, cuda, or cpu (default auto)
• conds (optional) - conds.pt for built-in voice references

Output: model (CHATTERBOX_VC), which feeds ChatterBox VC (Manual) - not the Simple variant, which expects the Auto loader's official API object.

Install and setup

Pack install via ComfyUI Manager (search "Latent Astronaut Suite") or:

cd ComfyUI/custom_nodes
git clone https://github.com/latentastronaut/comfyui-latent-astronaut-suite

Restart; install.py handles the chatterbox-tts package (installed --no-deps to dodge version conflicts) and librosa. Then drop the model file into ComfyUI/models/chatterbox/ and refresh. If the dropdown shows only "none," the file isn't there or the folder doesn't exist yet - that's the one real gotcha. Compared to TTS, this loader is refreshingly simple, so setup problems are almost always just the missing model file.