ElevenLabs is launching two speech transcription models: Scribe v3, for recordings, and Scribe v3 Live, for real-time. According to the company, Scribe v3 produces 27% fewer word errors than Scribe v2 across 102 languages, and Scribe v3 Live produces 43% fewer than Scribe v2 Realtime.
- ElevenLabs launches Scribe v3 for recordings and Scribe v3 Live for real-time transcription.
- Scribe v3 reduces word errors by 27% compared to Scribe v2 across 102 languages.
- Scribe v3 Live adds real-time speaker identification and cuts diarization errors by 39%.
- Eneco recognizes Dutch postal codes correctly on the first try 40% more often.
Transcription is the foundation of many audio products: meeting summaries, voice agents, dubbing, clinical documentation, or broadcast archives. An error made at this stage ripples through the entire chain, where it becomes far more costly to correct. Yet real-world audio is rarely clean. Multiple people speak at once, a customer hesitates mid-sentence, or a call involves two speakers not using their native language.
Demands That Shaped Scribe v3
Scribe has already been used to transcribe “hundreds of millions of hours” for note-taking apps, contact centers, clinical documentation tools, and voice agents. According to ElevenLabs, the teams using it requested five improvements. Meeting apps wanted better speaker identification in fast-paced exchanges and conversations with three or more people. Voice agent teams reported that a silence timer cuts off callers when they pause. Others requested an accuracy level closer to that of batch processing. Indian clients wanted Hindi as reliable as English, even when callers switch languages mid-sentence. Finally, many wanted more control over formatting, from writing numbers as digits to handling lists of terms long enough for a product catalog.
Scribe v3 makes 17% fewer diarization errors than Scribe v2 on recordings with three or more speakers, cases where, according to the company, most existing models fail. Diarization refers to identifying who is speaking at any given moment. Scribe v3 Live introduces this function in real-time and shows 39% fewer diarization errors than the best other real-time models tested.
For detecting speech turns, conventional systems rely on silences. A short threshold risks cutting off people who pause mid-sentence, while a long threshold creates a sense of slowness. Scribe v3 Live also relies on conversational context, so a caller who hesitates while giving a number is not cut off. This detection is already used in ElevenAgents and is now available via the API.
The model can also predict languages likely to appear in a recording and choose to display numbers as digits or words. “Keyterm prompting” accepts up to 1,000 expected terms, such as brand names, acronyms, or jargon, while the keyterm library, exclusive to Scribe v3 Live, contains thousands and only consults them when it hears a likely match. Formatting rules, written in plain English, can also be applied, for example to write every date in the YYYY-MM-DD format.
Results Measured Across Multiple Datasets
ElevenLabs compared its models to Scribe v2, Scribe v2 Realtime, and other popular models, on the same datasets. Multilingual accuracy was evaluated on an internal version of the FLEURS benchmark, which tests short, clean reading in 102 languages. This benchmark cannot measure diarization, as each sample contains only one speaker. The company therefore created its own dataset, composed of real meetings, calls, and voice notes in 16 languages, which it announces it plans to release as open source soon.
On FLEURS, Scribe v3 reduces the word error rate (WER) by 27% compared to Scribe v2, with improvement in every language tested. Gains are 42% in German, 35% in Spanish, and 23% in English, and reach 34% in Swahili, 29% in Tamil, and 28% in Thai. For Scribe v3 Live, ElevenLabs announces a 43% reduction in the semantic error rate, which only counts an error when the meaning changes, and a 31% reduction in WER compared to Scribe v2 Realtime. Greek improves by 77%, Danish by 54%, and Croatian by 45%.
In ElevenAgents, the Eneco Example
Eneco, which supplies energy to over 2 million households and businesses in the Netherlands, runs its voice agent on ElevenAgents and tested Scribe v3 Live before its launch. During real calls, the agent recognized Dutch postal codes correctly on the first try 40% more often, and house numbers 15% more often. Ruud Huigsloot, Senior Product Manager at Eneco, said: “ElevenLabs is a true partner for Eneco. With their engineers deployed on-site, we are working to further improve our voice agent.”
Scribe v3 and Scribe v3 Live run by default in the United States and are offered in isolated environments in the European Union, India, and Singapore, for businesses subject to data residency requirements. Zero Retention mode deletes audio and text immediately after a request ends. Both models are covered by SOC 2, ISO 27001, and PCI DSS Level 1 programs, and are HIPAA eligible. Existing integrations switch to Scribe v3 by replacing the model ID with scribe_v3 or scribe_v3_live. A version of Scribe v3 Live optimized for medical conversations is also offered in ElevenAgents.
