ElevenLabs has launched two new speech-to-text models: Scribe v3, for recorded audio, and Scribe v3 Live, a lower-latency variant for real-time audio. Min Kim set out the release in a 10 November blog post. Scribe v3 Live is available now in ElevenAgents, and both models are in ElevenAPI and are the default in the ElevenLabs app.
“Scribe v3 Live uses the context of the conversation as well as the audio, so a caller who pauses halfway through a number isn’t cut off.”
— Min Kim, in Introducing Scribe v3 and v3 Live
ElevenLabs says Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages, and that it had the most accurate speaker labels of any model the company tested. Scribe v3 Live, the post says, reduces semantic word error rate by 43% compared with Scribe v2 Realtime, and standard word error rate by 31%. It also adds speaker diarization, labelling who said what, to the company’s real-time model for the first time, plus improved turn detection.
Those figures come from ElevenLabs’ own evaluations. Multilingual accuracy was measured on an internal version of the FLEURS benchmark, which uses short, clean read speech in 102 languages. The company also built a 16-language set of real meetings, agent calls and voice notes for diarization, and says it will open-source that set. It scored each language on its own before averaging, and cautions that results measured in different ways aren’t directly comparable.
On FLEURS, the post says Scribe v3 improved in every language tested, including 42% fewer word errors in German, 35% in Spanish and 23% in English, with larger reductions in Swahili, Tamil and Thai. For the real-time model, the largest gains listed are in Greek, Danish, Croatian and Hindi. On recordings with three or more speakers, Scribe v3 is said to make 17% fewer diarization errors than Scribe v2. Scribe v3 Live is said to make 39% fewer diarization errors than the next-best real-time model ElevenLabs tested; on the English portion of the real-world set, its diarization error rate is given as 3.1%, against 5.9% to 7.3% for the other real-time models in that test. The company notes that those real-time diarization scores can’t be compared with the batch figures, because the transcripts were first aligned to a reference.
Kim writes that this is the same turn detection ElevenAgents already uses, now available in ElevenAPI. Other controls include keyterm prompting for up to 1,000 known terms, and transcript edits written in plain English, such as a rule to format dates. Scribe v3 Live adds a multilingual mode for conversations that switch language mid-sentence, and a keyterm library. Official documentation lists speaker diarization for up to 32 speakers on Scribe v3, and a keyterm library of up to 100,000 terms on Scribe v3 Live. Existing API integrations switch by changing the model ID to scribe_v3 or scribe_v3_live.
The model documentation quotes a 44% lower semantic word error rate for Scribe v3 Live on FLEURS across 73 languages, a different scope from the 102-language figures in the launch post. It also says the official SDKs don’t yet support turn prediction, multilingual mode or the keyterm library, and that those features need a direct WebSocket connection. Turn prediction, the docs say, has to be switched on with commit_strategy=turn_prediction; changing the model ID alone leaves manual committing in place.
The public API pricing table lists Scribe v3 at US$0.20 per audio hour and Scribe v3 Live at US$0.22. Optional keyterm prompting and transcript editing are listed as extra per-hour charges: US$0.050 and US$0.070 on the batch model, and, on Live, US$0.080 for prompting or the library and US$0.120 for editing. Those are transcription API rates, not an all-in ElevenAgents call price.
Eneco, which supplies energy in the Netherlands, tested Scribe v3 Live on its ElevenAgents voice agent ahead of the launch. In real customer calls, ElevenLabs says the agent recognised Dutch postcodes correctly on the first try 40% more often, and house numbers 15% more often. “It’s great to see significant results this early, and we are excited to continue collaborating,” said Ruud Huigsloot, a senior product manager at Eneco.
Both models run in the US by default and are offered in isolated EU, India and Singapore environments for enterprise customers with data-residency requirements. ElevenLabs says they are covered by its SOC 2, ISO 27001 and PCI DSS Level 1 programmes, are HIPAA-eligible with a Business Associate Agreement, and can be privately deployed. Zero Retention Mode, which deletes audio and text when a request completes, is available to select enterprise customers on eligible API traffic, not the web UI. A version of Scribe v3 Live optimised for medical conversations is also being released in ElevenAgents.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
