ElevenLabs Launches Scribe v3 and Scribe v3 Live: Lower Transcription Error Rates and Real-Time Speaker Diarization
ElevenLabs is replacing its transcription models with Scribe v3 for recorded audio and Scribe v3 Live for streaming. The company claims a 27% reduction in errors across 102 languages, real-time speaker attribution, and context-aware end-of-turn detection.
A flawed transcription is costly because the error cascades: into the meeting notes, into the summary, and ultimately into the action taken by an agent. That is the core premise behind ElevenLabs, which on November 10 launched Scribe v3 and Scribe v3 Live, its new speech-to-text models—one designed for recorded audio, the other for live streaming.
Who Said What, Even Live
The most noticeable change lies in diarization—assigning each sentence to the correct speaker. According to the company, Scribe v3 produces 17% fewer errors than its predecessor on recordings featuring three or more voices, where most models typically struggle. More importantly, Scribe v3 Live brings this capability to real time for the first time: achieving a 3.1% error rate on the English section of ElevenLabs’ test set, compared to 5.9% to 7.3% for other tested real-time models.
For voice agents, the other key addition is end-of-turn detection. Traditional systems listen for silence: if the pause threshold is too short, they cut off callers who hesitate; if it is too long, the agent feels sluggish. Scribe v3 Live factors in conversational context, ensuring that a caller who pauses mid-number sequence is not interrupted. Dutch energy provider Eneco, which tested the model on its phone agent, reported a 40% improvement in first-time postal code recognition.
In-House Figures Awaiting Independent Verification
ElevenLabs claims a 27% reduction in word errors compared to Scribe v2 across the 102 languages in its internal version of the FLEURS benchmark, with gains reported in every language, including 23% in English, 35% in Spanish, and 42% in German. In real time, “semantic” error—which only counts mistakes that alter the meaning—dropped by 43%. The model also accepts up to 1,000 key terms and formatting guidelines written in plain text (“format every date as YYYY-MM-DD”), while Scribe v3 Live handles conversations that switch languages mid-sentence.
These metrics, however, stem entirely from the developer: FLEURS consists of sentences read aloud under clean recording conditions, and ElevenLabs’ real-world dataset covering 16 languages is not yet public. The company promises to release it in the near future. No pricing details were disclosed in the announcement.
Our take
Transcription is fast becoming the critical layer for voice agents, and the battleground has shifted to real time: knowing who is speaking and when they have finished matters far more than squeezing out an extra point of accuracy on pristine audio. ElevenLabs is putting forward impressive figures, but they remain entirely self-reported; the promised release of its real-world test set will be the true test, especially as open-weights alternatives continue to advance, as seen with NVIDIA’s Nemotron 3.5 Transcribe. For European enterprises, isolated EU hosting and a zero-retention mode will carry just as much weight as raw precision.
Sources
- ElevenLabs — Introducing Scribe v3 and v3 Live, our most accurate speech to text models (November 10, 2026) Primary source
- ElevenLabs — Transcription (Speech to Text) documentation Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
