Scribe v3 / Scribe v3 Live

Scribe v3 and Scribe v3 Live are ElevenLabs' new Speech to Text models, announced on 10 November 2026, for recorded audio and realtime audio respectively. In a blog post by Min Kim, the company calls them its most accurate speech-to-text models yet, built for messy real-world audio: overlapping speakers, background noise, mid-sentence pauses, and language switching. ElevenLabs posted the announcement as @ElevenLabs.

According to the company, Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages on an internal version of FLEURS, with improvements in every language tested, and 17% fewer diarization errors than v2 on recordings with three or more speakers — the lowest diarization error of any model ElevenLabs says it tested. Scribe v3 Live is the low-latency realtime variant: it reduces semantic word error rate by 43% and word error rate by 31% versus Scribe v2 Realtime, adds speaker diarization to realtime transcription for the first time, and uses the whole conversation (not just a silence timer) for turn detection, so a caller who pauses while reading a number is less likely to be cut off.

Scribe v3 Live is available in ElevenAgents today. Both models are in ElevenAPI — existing integrations switch by changing the model ID to scribe_v3 or scribe_v3_live — and both are the default in the ElevenLabs app. The post also describes a Scribe v3 Live variant optimized for medical conversations in ElevenAgents. Default deployment is in the US, with isolated EU, India, and Singapore environments; Zero Retention Mode deletes audio and text when a request finishes. ElevenLabs says both models are covered by its SOC 2, ISO 27001, and PCI DSS Level 1 programs and are HIPAA-eligible with a BAA.

Highlights:

• Official blog (10 November 2026): https://elevenlabs.io/blog/scribe-v3 ;
• Official announcement: https://x.com/ElevenLabs/status/2120217467100541396 ;
• Product page: https://elevenlabs.io/speech-to-text ;
• These names belong to the Scribe speech-to-text line. They are not the same product as Eleven v4, ElevenLabs' text-to-speech model.

Why it matters

Transcripts sit at the start of meeting notes, voice agents, and clinical documentation: a missed digit or a mislabeled speaker is expensive downstream. Shipping a batch model and a realtime model under the same generation, with official figures for multilingual word error, diarization, and context-aware turn detection, is a concrete product step rather than a vague “more accurate” claim.

The percentages, the FLEURS setup, and the Eneco voice-agent example (first-try Dutch postcodes and house numbers) all come from ElevenLabs and a named customer. We have not independently verified them. For teams on Scribe v2, the useful objects are the official post, the model IDs, and the residency / retention terms.

Limits

Official scores use ElevenLabs' own evaluation sets and methodology; they may not be directly comparable with other vendors' published numbers. Medical optimization and HIPAA eligibility are as stated by the company — confirm contract and compliance terms before relying on them. Do not confuse Scribe v3 with Eleven v4 (text to speech).

1. Read the official post https://elevenlabs.io/blog/scribe-v3 and the announcement https://x.com/ElevenLabs/status/2120217467100541396 ;
2. In ElevenAPI, switch the model ID to scribe_v3 (recorded audio) or scribe_v3_live (realtime); Scribe v3 Live is also in ElevenAgents ;
3. Keep this distinct from Eleven v4 / Eleven v4 Turbo, which are text-to-speech models ;
4. For enterprise use, check official data-residency, Zero Retention, and HIPAA / BAA terms before switching production traffic.