ElevenLabs - Speech to Text

ElevenLabs isn't just a voice company - its Scribe model transcribes too, and this node wraps it. Feed it an AUDIO input and it returns the text, the detected language, and per-word timing. If you're building a pipeline that turns a recorded voice track into captions, a script for re-voicing, or a searchable transcript, this is the node that closes the loop inside ComfyUI.

You give it three required inputs: api_key, audio (any AUDIO from another node - a loaded file, a TTS result, a voice clone sample), and model (currently just scribe_v2). Then the controls that matter:

• diarize - label who's speaking. Turn it on and the transcript splits into speakers, which is the difference between "a wall of text" and "a usable podcast script."
• num_speakers - tell it how many voices to expect (0 = auto-detect). Setting this helps when diarization keeps merging two people.
• diarization_threshold - higher is stricter about separating speakers. If the diarizer is fragmenting one person into many, raise it.
• timestamps_granularity - word, character, or none. Choose word when you're going to rebuild captions from the timing.
• keyterms - comma-separated terms to bias recognition toward. The fix for Scribe consistently mangling your product name or a character's name.
• tag_audio_events - annotate sounds like (laughter) and (music) in the transcript. Nice for video editing notes.
• no_verbatim - strip filler words and stutters ("um", "uh") for a cleaner read.

Outputs: text (the transcript), language_code (e.g. en; empty input means auto-detect), and words_json (per-word timing and confidence, when you asked for it).

How to run it

Any node that produces AUDIO feeds it - a TTS node for round-tripping a script you want to verify, an audio loader for real recordings. For a voice-clone workflow, running STT on your reference samples is a good sanity check that the audio is clean enough to clone from. For captions, take words_json and the audio and drive a subtitle renderer with real timing instead of guessing.

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. Requires requests and soundfile - soundfile is what converts the AUDIO input to WAV for upload, so don't skip it.

Gotchas

• The pack re-executes API nodes every queue (IS_CHANGED), so each run is a fresh transcription - good for correctness, easy to forget when batching.
• Diarization isn't magic. Busy or overlapping speech still confuses it; setting num_speakers and tuning diarization_threshold is the actual fix, not re-running.
• For unlimited local transcription, whisper-based nodes are the free alternative; this one wins on the built-in diarization and event tagging being one node away.