ElevenLabs Speech to Speech

This is the node that makes dubbing look like magic. ElevenLabsSpeechToSpeech takes an audio recording of someone speaking, and re-speaks it in a different voice - keeping the content and the emotion while swapping the timbre. Laughs stay laughs, urgency stays urgency, it's just a different person now. It's voice conversion in the truest sense, and it's the difference between "good TTS" and "this actor never said that line."

Like every node in this group it's a partner node: the source audio goes up to ElevenLabs' hosted model through Comfy's API proxy, comes back as audio, and you're billed per minute from your Comfy account credits (Settings → Credits). No ElevenLabs account, nothing to install - it ships with ComfyUI core (landed February 2026).

The inputs that matter

• voice - the target voice you're converting into, wired in as an ELEVENLABS_VOICE from ElevenLabsVoiceSelector or ElevenLabsInstantVoiceClone. The dropdown-shaped hole in the workflow is a wire, not a dropdown - a recurring ElevenLabs-group lesson.
• audio - the source recording whose speech and emotion you're keeping.
• stability - same rule as TTS: low gives more emotional range, high gives steadier but flatter delivery. Default 0.5.
• model - eleven_multilingual_sts_v2 for non-English, eleven_english_sts_v2 for English-only. Picking one reveals speed, similarity_boost, and a style slider. Match the model to the language of the source for best results.
• remove_background_noise - default off. Turns on audio isolation on the source before conversion; reach for it when the source is a noisy field recording.
• output_format (mp3/opus) and seed - set and forget.

What it's actually for

Dubbing and ADR, where you want a performance rather than a flat re-read. Character work - take one clean read and restyle it into several distinct voices. Or the classic content trick: record once in your own voice with full emotion, then restyle into a polished narrator instead of re-recording until you sound like a pro.

Where people get burned

• Garbage in, garbage out. The model preserves emotion from the source. If the source read is flat, the converted take is flat - no amount of target voice fixes dead delivery.
• Noisy sources leak. Background noise rides along into the converted audio. Clean it with the toggle (or ElevenLabsAudioIsolation) before converting.
• Per-minute billing. Long sources are the expensive ones; trim before you convert.
• Language mismatch between the model and the source produces mangled output. If it's not English, you want the multilingual model.

Output is a single AUDIO into Preview Audio or Save Audio (Advanced), and then it's just another audio signal in your graph. Fast, persuasive, and occasionally eerie - which is about what you want from a voice-swap node.