ElevenLabs Instant Voice Clone

Want your voice - or your kid's, or a recording you legally own - coming out of a text-to-speech node? That's ElevenLabsInstantVoiceClone. You feed it 1 to 8 short recordings of a voice, it builds a clone on ElevenLabs' servers, and it hands back an ELEVENLABS_VOICE handle you can plug into any ElevenLabs speech node. Think of it as the "register a custom voice" step, turned into a graph node.

Like the rest of the ElevenLabs group it's a partner node: the samples are uploaded and processed remotely through Comfy's API proxy, billed once per clone from your Comfy account credits (Settings → Credits, no ElevenLabs key required). The node ships with ComfyUI core, so there's nothing to install.

The inputs

• files - an autogrow input holding between 1 and 8 audio recordings. More samples of varied sentences generally clone better, but even one clean take works. The node enforces the 1–8 range itself.
• remove_background_noise - default off. Flip it on and each sample gets run through ElevenLabs' audio-isolation backend before cloning. Turn this on if your recordings have hum or room tone; the clone quality is only as good as the cleanest sample you give it.

Where the voice goes

The output is an ELEVENLABS_VOICE, not audio. That's the point: you wire it into ElevenLabsTextToSpeech, ElevenLabsSpeechToSpeech, or ElevenLabsTextToDialogue, and from then on that voice is a first-class option in your workflow - no dropdown, just a wire. Because the clone is stored on your account, it persists across sessions.

The honest gotchas

• Your samples leave your machine. This is a cloud clone - the recordings are uploaded to ElevenLabs. Only clone voices you have the right to clone; consent and licensing are your problem, and they're real problems.
• One flat fee per clone regardless of sample count - so you might as well throw a few good takes in rather than one shaky one.
• Noise in = noise in the clone. A fan hum that's inaudible to you will be very audible in every sentence the clone speaks. Use the noise-removal toggle and re-record rather than hoping the model ignores it.
• There's no voice management here - no naming, deleting, or listing. This node just creates the handle; the API does the bookkeeping invisibly.

It's the node that turns "read this in my voice" from a promise into a wire. Just double-check the samples before you hit run - clones are sticky.