ElevenLabs has launched two new speech-to-text models: Scribe v3, for recorded audio, and Scribe v3 Live, a low-latency variant for realtime audio. In a 10 November 2026 blog post by Min Kim, the company says Scribe v3 “makes 27% fewer word errors than Scribe v2 across 102 languages” and that Scribe v3 Live “makes 43% fewer errors than Scribe v2 Realtime.”
Those percentages come from ElevenLabs’ own evaluations, not from an independent lab. The company’s model documentation describes Scribe v3 Live as having “44% lower semantic WER” than Scribe v2 Realtime on FLEURS. That documentation figure is not the one used here. The 43% comparison is the one given in the launch blog, which also sets out how the realtime test was scored.
Scribe v3 Live is available in ElevenAgents. Both models are now the default in the ElevenLabs app, and both are available in ElevenAPI under the model IDs scribe_v3 and scribe_v3_live.
What Scribe v3 and Scribe v3 Live add
Scribe v3 is the batch model, intended for recordings such as meetings, archives and clinical documentation. ElevenLabs says it has “the most accurate speaker labels of any model we tested.” Scribe v3 Live is the realtime model, and the company says it includes speaker diarization — labelling who said what — for the first time, together with improved turn detection.
On control of the output, the blog says keyterm prompting takes up to 1,000 terms the user already knows will appear, such as brand names, acronyms and jargon. Scribe v3 Live also adds a multilingual mode for conversations that switch language mid-sentence, and a keyterm library for larger vocabularies. The model documentation puts that library at up to 100,000 terms. Users can also set whether numbers come back as digits or words, and can apply formatting rules in plain English.
ElevenLabs says it is also releasing a version of Scribe v3 Live optimised for medical conversations, available in ElevenAgents. Both models, it says, run in the US by default and are offered in isolated EU, India and Singapore environments for enterprise customers with data-residency requirements. Zero Retention Mode deletes audio and text as soon as a request completes. The company says both models are covered by its SOC 2, ISO 27001 and PCI DSS Level 1 programmes, are HIPAA-eligible with a Business Associate Agreement, and are available for private deployment.
The launch sits beside other recent speech-model releases TSN has already covered, including ElevenLabs’ own Eleven Realtime speech-to-speech models for voice agents, Rime’s Prism speech-to-speech model, and Mistral’s Voxtral 3 speech models. Scribe v3 is a transcription launch, not a speech-to-speech model, but it is aimed at some of the same voice-agent and meeting-notes pipelines.
How ElevenLabs measured the new models
The 27% batch figure is a word-error comparison with Scribe v2 “across 102 languages.” Later in the same post, ElevenLabs says that on an internal version of the FLEURS benchmark — short, clean read speech in 102 languages — Scribe v3 “reduces WER by 27% compared with Scribe v2 and improves in every language tested.” It also reports semantic WER, which, it says, counts an error only when the meaning of the transcript changes. Language-level examples in that test include German improving by 42%, Spanish by 35% and English by 23%, with larger relative gains listed for Swahili, Tamil and Thai. Those are company-reported figures on its own internal FLEURS run.
Because every FLEURS clip has one speaker, the benchmark cannot measure diarization. ElevenLabs says it also built its own dataset of real meetings, agent calls and voice notes across 16 languages, and that it will open-source that set “soon.” It has not released it yet. On that work, the company says Scribe v3 has the lowest diarization error of any model it tested, with 12% fewer errors than Scribe v2 overall and 17% fewer on recordings with three or more speakers.
For realtime audio, the headline 43% figure is the one in the opening of the blog: Scribe v3 Live “makes 43% fewer errors than Scribe v2 Realtime.” A later section gives a more specific breakdown. Across 102 languages, ElevenLabs says, Scribe v3 Live “reduces semantic WER by 43% and WER by 31% compared with Scribe v2 Realtime,” improves in every language, and had a lower error rate than every other realtime model it tested. The method note matters: “we scored each language on its own before averaging, so all 102 languages count equally.” The company warns that many published results pool every clip into a single score, “so results measured in different ways aren’t directly comparable.”
On realtime diarization, ElevenLabs reports an English diarization error rate of 3.1% for Scribe v3 Live on its 16-language real-world dataset, compared with 5.9% to 7.3% for the other realtime models it tested. It says it aligned each model’s transcript to the reference before scoring, so that the measure reflects whether each word was assigned to the right speaker rather than whether words were missed. Those results, it says, “can’t be compared” with standard batch diarization scores. It also claims 39% fewer diarization errors than the next-best realtime model it tested.
A customer result is reported separately and should be read as such. Eneco, which ElevenLabs says supplies energy to more than two million homes and businesses in the Netherlands and already runs a voice agent on ElevenAgents, tested Scribe v3 Live on real customer calls ahead of the launch. In that test, the agent “recognized Dutch postcodes correctly on the first try 40% more often, and house numbers 15% more often.” That is a company-reported customer trial, not a public benchmark.
Company tests are not a substitute for an independent score
The useful part of ElevenLabs’ write-up is not the stack of percentages on their own. It is the way the company has tried, in the same post, to say what each number is. The 27% batch result is a word-error comparison with Scribe v2 across 102 languages on an internal FLEURS version. The 43% realtime result is the blog’s comparison with Scribe v2 Realtime, with a later, equal-weighted semantic-WER reading that matches that 43% figure. The 40% postcode improvement is Eneco’s result on its own Dutch calls. None of those is interchangeable with the others, and none is an outside audit.
That is also why the documentation’s “44% lower semantic WER” should not be silently swapped in for the blog’s 43%. The two pages disagree by a point on the same product, on the same day. For a launch story, the figure to use is the one the company put in the announcement and then explained, not the slightly different number on the model card.
There is a further gap. The 16-language real-world set that underpins the diarization claims is still only promised. Until it is actually released, those speaker-label scores cannot be re-run by anyone else. FLEURS, even in an internal version, is read speech, not overlapping meetings. A contact-centre or clinical team that wants to know whether Scribe v3 will cut errors on its own audio still has to test that audio. The Eneco postcode result is encouraging for that kind of work, but it is one customer, on one task, reported by the vendor.
Scribe v3 and Scribe v3 Live are available now, and the upgrade path ElevenLabs describes is a change of model ID. The claims that come with them are the company’s. They should be repeated as the company’s, and checked against the job the transcript has to do.
Sources
- ElevenLabs, “Introducing Scribe v3 and v3 Live, our most accurate speech to text models”, Min Kim, 10 November 2026. https://elevenlabs.io/blog/scribe-v3
- ElevenLabs, “Models” documentation (Scribe v3 and Scribe v3 Live). https://elevenlabs.io/docs/overview/models
Related stories
ElevenLabs’ earlier Eleven Realtime speech-to-speech models were a voice-agent launch from the same company; Scribe v3 is a transcription model aimed at some of the same pipelines.
Also relevant:
- Rime Launches Prism Speech-to-Speech Model With Inspectable Text
- Mistral Launches Voxtral 3 Speech Models and a Voice Agent SDK
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
