ElevenLabs - Speech to Speech

Text to Speech starts from text. Speech to Speech starts from audio: it takes an existing recording, keeps the delivery - the pacing, the emotion, the pauses - and re-voices it in a different voice. This is how you'd make a character "speak" your own performance, or redo a narration in a new voice without recording it again. The node hits the ElevenLabs /speech-to-speech/{voice_id} endpoint and returns the re-voiced audio.

Three required inputs: api_key, audio (the source recording, as an AUDIO dict), and voice_id - the target voice, from the selector or fetch nodes. Then model: eleven_multilingual_sts_v2 (default) or eleven_english_sts_v2. Multilingual if the source audio isn't English, English-only otherwise.

The tuning knobs will look familiar if you've used the TTS node:

• stability - low for expressive, high for consistent.
• similarity_boost - how closely it matches the target voice.
• style - expression exaggeration; keep at 0 unless you want a performance.
• speed - nudge the delivery faster or slower (0.5–2.0).
• remove_background_noise - clean the source before conversion. Leave it on for phone recordings or room-mic captures; it's cheap insurance.
• seed and output_format - the usual reproducibility and format controls, with the Creator-tier note on the high-bitrate formats.

Single output: audio (AUDIO).

When you'd reach for it

The classic workflow is performance transfer: record yourself reading a line with the exact emotion you want, then STS it into a cloned or premade character voice. That beats prompting TTS for "angry whisper" every time, because the acting is in your recording, not in the prompt. It's also how people re-voice existing clips - an old narration, a legacy recording - into a consistent brand voice, or dub a voice track into another character in the same script. If you're doing dialogue, you can STS each actor's raw take into their assigned voice and keep the natural back-and-forth timing.

Installing it

Part of the ComfyUI API Toolkit pack. Manager: search "API Toolkit". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-API-Toolkit
cd ComfyUI-API-Toolkit
pip install -r requirements.txt

Restart. Needs requests and soundfile (soundfile converts your AUDIO input to uploadable WAV).

Gotchas

• The result is only as clean as the source. If the input has echo or music behind it, the conversion carries that baggage into the output voice. Use remove_background_noise, or pre-clean with AIS_EL_AudioIsolation first.
• Per-second or per-character billing applies - long source audio gets expensive fast, and each queue re-runs because the pack uses IS_CHANGED.
• This is one of the "data leaves the machine" nodes, and voice re-voicing is exactly the capability that makes nonconsensual cloning a real concern. Only feed it audio you have the right to re-voice.