ChatterBox VC (Manual)

ChatterBox VC (Manual) is voice conversion: you give it a clip of someone talking and a reference clip of the voice you want, and it re-synthesizes the speech in that target voice. Same text, same words, new timbre. This is the "dub your character with a different voice" node, and it's surprisingly good at it. If you've seen ChatterBox TTS floating around ComfyUI - it got real community traction through 2025 for voice cloning - this is the sibling that swaps voices instead of reading text.

How it works

The mechanism is compact. Both your audio (the clip to convert) and target_voice (the reference) get written to temp WAV files. The target voice is fed through the S3Gen model's reference encoder to build a voice embedding, while the input audio is tokenized into speech tokens. Then S3Gen re-synthesizes those tokens conditioned on the target voice. One model does the whole job - which is why the loader for this node only needs s3gen.safetensors (plus optional conds.pt for built-in voices), not the whole file set that TTS requires.

Inputs and outputs

Only three inputs, all required:

• model - a CHATTERBOX_VC model from a loader (ChatterBox VC Loader Manual or Auto)
• audio - the speech you want converted
• target_voice - reference audio; the tooltip says 6–10 seconds is ideal, and it's not kidding. Short references mean a weak voice embedding.

One output: audio, which wires into any ComfyUI audio consumer - a preview node, VHS video save, whatever your pipeline uses.

Getting it running

Install the pack via ComfyUI Manager (search "Latent Astronaut Suite") or:

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

Restart ComfyUI. On first boot, install.py installs the chatterbox-tts package with --no-deps (its strict version pins fight ComfyUI's environment) plus librosa. The Manual loader reads model files from ComfyUI/models/chatterbox/; the Auto loader downloads from HuggingFace and caches.

Where people get burned

First run is slow - the model loads and the download happens once, so don't judge it on iteration one. The pack disables ChatterBox's built-in watermarker because it segfaults on version mismatches; that's intentional, not a broken install. If audio comes back empty or crashes, check your reference length (trim to ~8 seconds) and make sure both clips are actual speech, not silence. And be aware this is a heavy model - give it GPU headroom or it'll swap to a slow CPU fallback on its own.