ElevenLabs launches Scribe v3 with 27% fewer word errors across 102 languages, plus realtime Scribe v3 Live

ElevenLabs has launched two speech to text models: Scribe v3 for recorded audio, which the company claims makes 27% fewer word errors than Scribe v2 across 102 languages, and Scribe v3 Live, a low latency variant for realtime audio that it says makes 43% fewer errors than Scribe v2 Realtime. Pricing is not yet disclosed.

Availability

Scribe v3 Live is available today in ElevenAgents. Both models are available in ElevenAPI, where ElevenLabs says switching from Scribe v2 is a one-line change to the model ID. ElevenLabs says Scribe has powered hundreds of millions of hours of transcription for meeting-notes apps, contact centers, clinical documentation tools and voice agents.

What changed

Scribe v3 (batch)

• 27% lower word error rate (WER) than Scribe v2 on FLEURS across 102 languages, with improvement in every language tested, according to ElevenLabs.
• Per-language gains: German 42%, Spanish 35%, English 23%, Swahili 34%, Tamil 29%, Thai 28%.
• 12% fewer diarization errors than Scribe v2 overall and 17% fewer on recordings with three or more speakers. ElevenLabs says this is the lowest diarization error of any model it tested.
• Language hints inferred from the recording, and a choice of numbers returned as digits or words.
• Keyterm prompting accepts up to 1,000 terms such as brand names, acronyms and jargon. Transcript edits apply formatting rules written in plain English, for example "write every date as YYYY-MM-DD."

Scribe v3 Live (realtime)

• 43% lower semantic WER and 31% lower WER than Scribe v2 Realtime across 102 languages, with improvement in every language, per ElevenLabs.
• Largest gains: Greek 77%, Danish 54%, Croatian 45%, Hindi 36%.
• Speaker diarization in realtime for the first time: 39% fewer diarization errors than the next-best realtime model tested. On the English portion of ElevenLabs' own dataset, the diarization error rate is 3.1%, against 5.9% to 7.3% for other realtime models.
• Turn detection that uses conversation context as well as audio, rather than silence alone. This is the same turn detection ElevenAgents already uses, now exposed in ElevenAPI.
• A multilingual mode for conversations that switch languages mid-sentence, and a keyterm library for vocabularies too large for a list. The library holds thousands of terms that might appear, and the model consults it only when it hears a likely match.

How ElevenLabs measured it

All results come from ElevenLabs' own evaluation against Scribe v2, Scribe v2 Realtime and other speech to text models, scored on the same datasets. Multilingual accuracy used an internal version of the FLEURS benchmark, which covers short, clean read speech in 102 languages. Because FLEURS clips have one speaker, ElevenLabs built its own diarization dataset of real meetings, agent calls and voice notes across 16 languages. It says it will open source the dataset soon.

Scores were computed per language and then averaged, so all 102 languages count equally. ElevenLabs notes that many published results pool all clips into one score, so results measured differently are not directly comparable. For realtime models it leads with semantic WER, which counts an error only when the meaning changes. The realtime diarization figures were scored after aligning each transcript to the reference, so they cannot be compared with the batch diarization results.

ElevenLabs also says Scribe v3 makes fewer word errors than every other batch (non-streaming) model it tested, compared on the languages both models support. Scribe v3 Live has a lower error rate than every other realtime model it tested.

Customer testing

Eneco, which supplies energy to more than 2 million homes and businesses in the Netherlands, runs its voice agent on ElevenAgents and tested Scribe v3 Live on it.

What this means

The release targets problems ElevenLabs says customers raised: diarization with three or more speakers, callers cut off mid-sentence, the accuracy gap between realtime and batch, Hindi and English code-switching, and output control. Realtime diarization and context-based turn detection in ElevenAPI move features previously tied to ElevenAgents into the standalone API. All accuracy figures are vendor-reported on ElevenLabs' internal FLEURS variant and its own 16-language dataset, which has not yet been released, so independent verification awaits that open-sourcing.