Scribe v3 by ElevenLabs - Speech to text with 27% fewer errors in 90+ languages
by•
Scribe v3 is ElevenLabs’ most accurate speech-to-text model, for developers building on calls, meetings, interviews and media. It makes 27% fewer word errors than Scribe v2 across 90+ languages. Unlike transcription that breaks on noise, accents and crosstalk, it labels up to 32 speakers, tags laughter, music and emotion, boosts up to 1,000 of your own terms, detects 56 entity types and handles files up to 15 hours. For live use, Scribe v3 Live runs under 150ms.
Replies
Hunting Scribe v3 today, the new speech-to-text model from ElevenLabs.
A few numbers that stood out. On their benchmarks it lands at 5.5% WER vs 9.7% for GPT-Transcribe. Fieldy says user retention went up 50% after switching to Scribe. And it’s already running at Meta and Twilio.
If you’re testing it, skip the clean studio audio. Use your messiest call recording, add your product names as keyterms, and turn on diarization.
Works in the dashboard and via API, no sales call needed. Go test it https://elevenlabs.io/blog/scribe-v3
@kate_ramakaieva congrats on the hunt Kate Great work bringing scribe v3 to the community.
The 1,000 custom keyterms is the detail that caught my eye — most transcripts I deal with fall apart on product names and internal jargon, not clean speech. Same with the 32-speaker diarization and the laughter, music, and emotion tags; those are the bits that usually get lost. Curious how well the keyterm boost holds up on messy, overlapping call audio versus the clean benchmarks.
What's the best way to tackle this when there are names that aren't comprehensive by the software?
labelling up to 32 speakers in messy calls is a big deal. congrats elevenlabs!
Boosting a thousand of my own product names as keyterms is the fix for the thing that makes every transcript useless, since no general model has ever heard of the internal tool names we say in standups.