ElevenLabs Voice Isolator

Some of the best ComfyUI workflows start with garbage audio: a phone recording of a lecture, a clip with a fan humming in the background, an interview where two people talk over traffic. This is the cleanup node. Feed it any audio and it isolates the voice, dropping as much background noise as ElevenLabs' model can manage.

How it works

It uploads your audio tensor as a WAV to ElevenLabs' /v1/audio-isolation endpoint and returns the cleaned result as a fresh AUDIO tensor. Two inputs, one output, no dials to fight with - this is the simplest node in the pack, and it's meant to be. The model is doing the whole job; your only input choices are the api_key and the audio itself.

That simplicity is also its ceiling. There's no "aggressiveness" or "preserve music" slider here - you get the default isolation profile and that's it. If the output sounds over-processed or the voice gets thinned out, you can't tune it from this node. You either accept it or reach for a local tool with more control.

What it pairs with

This node earns its place in the middle of a pipeline, not at the end:

• Clean a recording, then feed it to Voice Changer for re-voicing without the background mush getting re-synthesized into the target voice.
• Clean dialogue, then run it through the pack's Speech-to-Text node (not covered here, but it's in the same install) for transcription.
• Recover a voice track from a noisy clip before dubbing, so the dubber isn't fighting the hum.

One nice robustness detail worth knowing: on a failed API call the node returns your original audio unchanged rather than an empty tensor. Your graph doesn't silently break with silence - it degrades to the input, and the error is logged to the console. That's a small thing, but it's the difference between noticing and not noticing a failed run.

Install

Same as the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/karthikg-09/Elevenlabs-ComfyUI.git
cd ElevenLabs-ComfyUI && pip install -r requirements.txt

or ComfyUI Manager → search "Elevenlabs-ComfyUI" → restart. Add Node → ElevenLabs.

The honest take

For a quick "make this recording listenable" job, this is the fastest path in the pack: no model download, no dependency stack, one API call. The open ecosystem has strong local alternatives - Demucs and similar source-separation models are free and good - but they cost setup time and VRAM, which this node spends exactly zero of. Where it loses is on long-form batch cleanup: every clip is a separate metered call, and if you're cleaning hours of material you'll notice the credit burn. Short clip, one-off cleanup, want it now? This node. Regular high-volume denoising? Stand up something local and keep the API budget for the jobs only ElevenLabs can do.