ElevenLabs Scribe v3 Cuts Transcription Errors by 27 Percent
By Ganesh Karki
Key facts
- Section
- Tech & AI
- Country
- USA
- Published
- 18 Nov 2026
ElevenLabs Scribe v3 is here: the new speech-to-text model makes 27 percent fewer word errors than Scribe v2 across 102 languages, at the same hourly price. ElevenLabs released Scribe v3 and Scribe v3 Live on 10 November 2026, and both are now the default transcription models in the ElevenLabs app. The Live version transcribes speech as it happens and, for the first time in the company's realtime line, labels who is speaking.
ElevenLabs Scribe v3 at a glance
| Detail | Information |
|---|---|
| Released | 10 November 2026 |
| Models | Scribe v3 for recordings, Scribe v3 Live for realtime |
| Word errors | 27% fewer than Scribe v2, across 102 languages |
| Live errors | 43% fewer than Scribe v2 Realtime |
| Speaker labels | 12% fewer errors, 17% fewer with three or more speakers |
| API price | $0.20 per hour for v3, $0.22 per hour for Live |
| Model IDs | scribe_v3 and scribe_v3_live |
What happened
Scribe v3 is built for recorded audio such as meetings, interviews, podcasts and voice notes, handling files up to 15 hours long through the API. Scribe v3 Live works in realtime, for example on a phone line or inside ElevenAgents, the company's voice-agent builder, and it predicts when a speaker has finished, using the conversation as well as the sound. Developers switch from v2 by changing one model name.
The headline numbers come from ElevenLabs itself. DataNorth reports the 27 percent figure was measured on the company's own version of FLEURS, a public multilingual speech benchmark, plus its own set of meetings, agent calls and voice notes in 16 languages. Creative AI News notes the API price did not move from Scribe v2, and that a medical version of Scribe v3 Live is coming through ElevenAgents. The company says it will open-source its diarisation test set.
Why it matters
Transcription is the unglamorous plumbing under voice agents, meeting notes and captions, and a 27 percent error cut at the same price is a genuine improvement for everyone who relies on it. Speaker labels in realtime matter most: a live call transcript with names attached is far more useful than a wall of unattributed text. The release also lands as voice agents spread into healthcare and customer service, the same shift behind ChatGPT healthcare agents and the self-hosted Claude managed agents now in private beta.
What is not confirmed
Every accuracy number so far is ElevenLabs measuring itself, and there is no independent check yet. Artificial Analysis, the leaderboard most transcription launches quote, still lists only Scribe v2 at a 2.2 percent word error rate, fourth on its board. The German figure of 42 percent fewer errors comes from short, clean read speech, not noisy interviews. Inside the app, transcription is paid in credits rather than at the 0.20 dollar API rate, so the headline price does not describe every user.
Sources
Sources and references
Information checked: 18 November 2026
Related stories
SpaceX Falcon 9 NROL-136 Launch: What to Watch on 21 November
SpaceX Falcon 9 NROL-136 launch is scheduled for 21 November from Vandenberg, carrying reconnaissance satellites for the National Reconnaissance Office.
20 November 2026AO3 Returns After 17 Days Offline After Security Breach
AO3 returns after 17 days offline following a security incident, with some features still disabled, and went down again briefly on 19 November for maintenance.
20 November 2026ChatGPT for Accounting Launches With QuickBooks Link
ChatGPT for Accounting launched this week, connecting the chatbot to QuickBooks, Xero and NetSuite so finance teams can reconcile books and close the month.
20 November 2026Tesla Cybercab Rides Begin in Dallas With V1.5 Cars
Tesla Cybercab rides began in Dallas on 19 November, the second city after Austin to get the steering wheel-free robotaxi, with upgraded V1.5 vehicles.