Voice Replace

Voice Replace is the sibling node to Voice Clone in this pack, and the difference matters more than people expect. Voice Clone makes a voice say new text. Voice Replace takes an existing recording and swaps the speaker - the words, the pacing, the pauses stay, but the timbre becomes whoever you hand it as the reference. That's voice conversion (VC), not TTS, and it's the tool you reach for when you want to redub a line, replace a narrator in a track you already have, or make a character's existing dialogue come out in another character's voice.

Inputs and output

The node is deliberately bare-bones - three inputs, one output, no sampler knobs at all:

• audio - the recording you want to re-voice. Can come from anywhere in ComfyUI: another TTS node, an audio file node, even the audio track of a generated video.
• target_voice - a reference clip of the voice you want instead. A few seconds of clean, single-speaker audio works best.
• disable_watermark - set true to skip the PerTh watermark (the same one Voice Clone applies by default).

Output is a single AUDIO tensor at 24 kHz, so it plugs into the same downstream nodes as Voice Clone's output.

How it works

Under the hood it's Chatterbox's VC model (ChatterboxVC). The node takes both clips, converts them to mono, resamples them to the model's rate, writes them to temp files, and runs the conversion. A nice detail in the source: it handles whatever you throw at it - stereo gets averaged to mono, mismatched sample rates get resampled automatically. That makes it forgiving to wire into sloppy workflows where your audio came from five different sources.

One thing to know going in: this is not a transcription-and-resynthesis pipeline. It's converting the existing speech in place, so the content of your audio is what you hear in the output - just with a different voice on top. If you wanted the target voice saying different words, that's Voice Clone's job, not this node's.

Install and setup

Same pack, same install as Voice Clone: search "ComfyUI Voice Clone" in ComfyUI Manager, or

cd ComfyUI/custom_nodes
git clone https://github.com/Sean-Bradley/ComfyUI-Voice-Clone.git

then pip install -r requirements.txt and restart. The model files go in ComfyUI/models/tts/chatterbox/ - five files (conds.pt, s3gen.safetensors, t3_cfg.safetensors, tokenizer.json, ve.safetensors) from the pack's Hugging Face folder.

Gotchas

• It's slow to start. Like Voice Clone, the VC model is loaded fresh on every run, so each generation pays a load cost before it does anything. Don't assume it's hung on the first attempt.
• Feed it a clean target voice. A reference with background music, a second speaker, or heavy reverb will leak into the conversion.
• Quality ceiling. Chatterbox's VC is good, but it's not a Hollywood dubbing tool - heavily accented or overlapping speech can come out garbled. Keep source audio clear and you'll get the best results.
• The error path in the source is commented out, so when it fails it fails loud rather than returning silence. If you see an exception, the clip that broke it is usually the culprit - swap the reference.