DIGIT ElevenLabs Speech to Speech

Most TTS starts from text. DIGIT ElevenLabs Speech to Speech starts from audio: you feed it a voice recording and a target voice ID, and it re-speaks the words in the new voice while keeping the original delivery - the pacing, the emphasis, the emotional read. That distinction is the whole node. You're not re-synthesizing the line from text; you're re-voicing an existing performance.

The use cases are the obvious ones: you recorded a scratch VO yourself and want a polished voice on top, you have a decent AI voice that's one step away from the client's preferred voice, or you need to swap a voice across an existing track without rewriting the script. It's the "fix it in post" node of the voice world.

How it works

You connect audio (an AUDIO tensor - from the pack's ElevenLabs TTS/STT nodes, or loaded from a file) and give it voice_id, the target voice. The model choice is eleven_multilingual_sts_v2 (default) or eleven_english_sts_v2 if you're working in English and want the single-language path. Then the delivery knobs:

• stability (0.5) - lower means the model is more willing to get expressive; higher means more consistent.
• similarity_boost (0.75) - how closely it matches the target voice. Crank it for a stricter clone.
• speed (0.7–1.3), style (0–0.2), use_speaker_boost, and remove_background_noise for cleaning up a messy source take.
• seed for reproducible runs; output_format (pcm_44100 default, or MP3/Opus).

api_key is optional - it auto-detects ELEVENLABS_API_KEY or the pack's DIGIT_ELEVENLABS_API_KEY. Output is a single audio tensor.

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/thedepartmentofexternalservices/comfyui-digit.git
cd comfyui-digit
pip install -r requirements.txt

Or ComfyUI Manager → search comfyui-digit → install → restart. Then:

export ELEVENLABS_API_KEY=your_key_here

What trips people up

The mental model people get wrong: STS is not TTS-with-a-reference. It needs actual audio in, and it's preserving that audio's delivery - so if your source take is flat and robotic, the output will be a flat, robotic read in a different voice. Garbage in, garbage out, but the garbage now has a better voice. Also, similarity_boost interacts with stability: pushing both to extremes can make the output feel a bit processed, so the 0.75/0.5 defaults are a decent starting point. And like everything ElevenLabs, it's per-call billed, so treat the seed as your friend when you're iterating - same seed, same output, no wasted retries.