ElevenLabs has released Scribe v3 for recorded audio and Scribe v3 Live for real-time transcription, claiming 27% fewer word errors across 102 languages on its own tests. Why it matters: the Live model is the first Scribe that can label who is speaking while the conversation is still happening — the feature voice agents and meeting tools have been missing.
November 10, 2026 — Both models are available now through ElevenAPI, Scribe v3 Live is live inside ElevenAgents, and Scribe v3 has become the default transcriber in the ElevenLabs app. Existing integrations switch over with a model ID: scribe_v3 or scribe_v3_live.
The accuracy claims — with the label on
Every number in this launch comes from ElevenLabs’ internal evaluations, not a third-party benchmark, and the post flags them that way throughout. On an internal run of the FLEURS dataset across all 102 languages (October 23–29), Scribe v3’s word error rate fell from 17.0% with Scribe v2 to 12.4%, while semantic word error — mistakes that actually change meaning — fell from 5.4% to 4.4%. The company reports per-language word-error drops including German at 42%, Spanish at 35%, and English at 23%, with every tested language improving.
Scribe v3 Live posts bigger relative gains against Scribe v2 Realtime on the same dataset: word error down from 26.5% to 18.2%, and semantic word error down 43%, from 12.5% to 7.1%. Streaming models are judged on the semantic figure because they emit text before hearing the full sentence.
Speaker labels, turn-taking, and vocabularies
The structural news is diarization. On ElevenLabs’ own 16-language set of real meetings, agent calls, and voice notes, the batch model’s diarization error improved to 5.6% from 6.4%. The Live model adds speaker labels to real-time transcription for the first time: 3.1% diarization error on English real-world recordings in the company’s aligned comparison, against 5.9% for Grok Voice Transcribe 2.0 — again, the company’s own scoring, with the underlying real-world set promised for open release.
Live also borrows the context-based turn detection already used in ElevenAgents, so an agent is less likely to talk over a caller pausing mid-sentence, and it can switch languages mid-sentence in a multilingual mode. Both models accept keyterm prompts — up to 1,000 terms for the batch model, a 10,000-term library for Live — and the batch model accepts plain-English editing rules for output formatting. ElevenLabs cites one pre-launch customer result, also company-reported: Dutch energy supplier Eneco saw first-try postcode recognition rise 40% and house numbers 15% on real calls.
Why it matters
Transcription is quietly becoming the load-bearing layer under voice agents: one wrong name, number, or address in the transcript becomes a wrong summary, a wrong action, a wrong bill. A 27% error cut, if it holds on messy real-world audio rather than clean test sets, is the kind of unglamorous improvement that decides whether agents survive production. The limits are worth stating plainly — the docs list a 5 GB file cap, a 15-hour duration limit, and diarization for up to 32 speakers, and none of the headline numbers have been independently replicated yet.
Frequently asked questions
What is new in ElevenLabs Scribe v3?
Two models: Scribe v3 for recorded audio and Scribe v3 Live for real-time streams. ElevenLabs reports 27% lower word error across 102 languages for the batch model, plus better diarization, keyterm support, and plain-English formatting rules.
Can Scribe v3 Live tell speakers apart in real time?
Yes — it is the first Scribe model with real-time speaker diarization. ElevenLabs reports 3.1% diarization error on English real-world recordings in its own comparison; independent replication is still pending.
Where can you use Scribe v3?
Through ElevenAPI, inside ElevenAgents (Live), and as the default transcriber in the ElevenLabs app, using the model IDs scribe_v3 and scribe_v3_live.
Sources: ElevenLabs launch coverage and documentation via Nvdium, RuntimeWire