ElevenLabs Scribe v3 Live cuts transcription errors 43%, adds speaker ID

Summary

• ElevenLabs launched Scribe v3 and Scribe v3 Live, its two new speech-to-text models, across 102 languages.
• Scribe v3 Live makes 43% fewer errors than v2 Realtime and identifies speakers in live audio.
• Call centers and voice agents gain transcription that labels speakers and avoids cutting off pausing callers.

The latest

Two new speech-to-text models from ElevenLabs promise sharper live transcription for voice agents, the company said in a November 10, 2026 blog post by Min Kim. Scribe v3 Live makes 43% fewer errors than Scribe v2 Realtime on semantic word error rate, while the recorded-audio model Scribe v3 cuts word errors 27% against its predecessor across 102 languages on the FLEURS benchmark.

Details

• The models: According to the company, Scribe v3 handles recorded audio and Scribe v3 Live handles streaming, replacing Scribe v2 and Scribe v2 Realtime. They appear in the API as scribe_v3 and scribe_v3_live, and both are now the default option inside the ElevenLabs app.
• The error numbers: ElevenLabs said Scribe v3 Live also posts a 31% lower raw word error rate alongside the 43% semantic gain. On the recorded model, it reported improvements of 42% in German, 35% in Spanish, 34% in Swahili, 29% in Tamil, 28% in Thai and 23% in English.
• Live gains by language: The biggest streaming improvements listed were Greek at 77%, Danish at 54%, Croatian at 45% and Hindi at 36%. ElevenLabs did not publish latency figures or pricing for either model, and made no mention of Arabic or dialect support.
• Speaker separation: Scribe v3 makes 12% fewer speaker-identification errors than v2 overall, rising to 17% on recordings with three or more speakers, per the company. Scribe v3 Live brings speaker labeling to real-time transcription for the first time, at a 3.1% error rate in English.
• Against rivals: ElevenLabs said competing real-time models it tested scored between 5.9% and 7.3% on English speaker-identification error, putting Scribe v3 Live 39% ahead of the next-best streaming model. It did not name the competing systems included in the comparison.
• Turn prediction: Scribe v3 Live reads conversational context alongside audio to judge when a speaker has finished, so a caller who pauses mid-sentence is not cut off. It is the same turn detector used in ElevenAgents, now exposed through the API via a commitStrategy setting.
• Language controls: Both models cover 102 languages. Scribe v3 infers likely languages and can render numbers as digits or words, with keyterm prompting accepting up to 1,000 terms. The live model adds a multilingual mode for mid-sentence language switching and a keyterm library holding thousands of entries.
• Availability: Scribe v3 Live is live in ElevenAgents now, and both models are available through the API. A version of Scribe v3 Live tuned for medical conversations is rolling out to healthcare teams in ElevenAgents, without a date given for general release.
• Enterprise compliance: Both models run in the United States by default, with isolated environments in the European Union, India and Singapore for data residency. A Zero Retention Mode deletes audio and text once a request completes, and the models fall under SOC 2, ISO 27001 and PCI DSS Level 1 programs, with HIPAA eligibility and private deployment offered.
• Customer result: On real calls, the voice agent run by Dutch energy firm Eneco recognized Dutch postal codes correctly on the first attempt 40% more often and house numbers 15% more often, ElevenLabs said. Rod Hooijschuur, a senior product manager at Eneco, said it was "great to see significant results this early."

Background

Speech-to-text accuracy is measured by word error rate, the share of words a model gets wrong. Semantic word error rate weights those mistakes by how much they change meaning, so a model can score better by failing on words that matter less.

Between the lines

The pairing of real-time speaker labeling with turn prediction points at contact centers rather than media transcription: both features matter most when a machine must decide who is talking and when to reply. Isolated environments in three regions, Zero Retention Mode and HIPAA eligibility point the same way, toward regulated buyers in healthcare and finance who cannot move call audio across borders.

What's next

Watch for the medical-conversation version of Scribe v3 Live to reach healthcare teams in ElevenAgents, and for ElevenLabs to publish latency figures, pricing, or the identity of the competing real-time models in its speaker-identification comparison.

Source: ElevenLabs