ChatterBox VC (Simple)

ChatterBox VC (Simple) is voice conversion at its most minimal: three inputs, one output, no sampler knobs to touch. You hand it a model from the Auto loader, the audio you want converted, and a reference clip of the target voice - it re-synthesizes your clip in that voice and returns the audio. Where the Manual VC node does the conversion by hand, this one calls ChatterBox's official generate() API and lets library defaults handle everything else. It's the "I just want the voice swapped" option.

How it works

The node writes both input clips to temp WAV files, then calls model.generate(audio=..., target_voice_path=...) on the official ChatterBox VC model object. Internally that's the same two-step S3Gen flow - the target voice gets embedded as a reference, your audio gets tokenized, and the tokens are re-synthesized under that voice - but you never see the mechanics. Output audio is packaged as a ComfyUI AUDIO object at the model's sample rate, and the temp files are cleaned up after each run.

One important pairing detail: model here expects the Auto loader's official model object (the tooltip says "from auto-loader"). The Manual VC node uses the pack's own hand-built wrapper; this node uses ChatterBox's API. Don't cross the streams.

Inputs and outputs

• model - a CHATTERBOX_VC model from ChatterBox VC Loader (Auto)
• audio - the speech to convert
• target_voice - the reference voice, 6–10 seconds ideal

Output: audio - ready for any ComfyUI audio consumer.

That's the whole surface. No exaggeration dial, no temperature, no seed - if you want per-run consistency you'll want the Manual node, which also has those dials.

Install

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. First boot runs install.py, installing chatterbox-tts with --no-deps plus librosa. The model itself downloads via the Auto loader on first use, so the first run is slow and subsequent ones are fast.

Where people get burned

The loader-matching mistake is the recurring one - Manual model into Simple node, or vice versa, and things just break. Also, since there's no seed, output varies run to run; if you're converting a long clip for a video, budget for that. And give it a clean reference: the 6–10 second guidance is there for a reason, and background music in the target voice clip will bleed into the result.