Loading page…
The release targets interrupted callers, crowded meetings and specialized vocabulary. Its accuracy claims rest on internal evaluations using different measures of transcription errors.
Listen to this story
ElevenLabs’ November 10 release makes Scribe v3 available for recorded audio and Scribe v3 Live for streaming through its API and app. The live model adds speaker labeling and uses conversation context to judge when a caller has finished, giving voice-agent builders an alternative to fixed silence timers. The company reports accuracy gains from internal tests, but they use different metrics and datasets; developers can upgrade by changing model identifiers, and enterprise customers have regional and private deployment options.
Internal tests found relative reductions of 27% in standard word error for Scribe v3, and 43% in semantic and 31% in standard word error for Live, versus their predecessors.
For recordings with three or more speakers, ElevenLabs reports 17% fewer labeling errors than Scribe v2; it plans to open-source the separate 16-language test set.
Eneco reported 40% more first-try Dutch postcode recognition and 15% more house-number recognition, but no underlying success rates were disclosed.
Live transcripts can now identify who said what in ElevenLabs’ latest speech-recognition release, while voice agents get more context for deciding when a caller has finished speaking. The company released Scribe v3 for recorded audio and Scribe v3 Live for streaming transcription on November 10, 2026, alongside company-reported accuracy gains across 102 languages.
Both models are available through ElevenAPI and are defaults in the ElevenLabs app. Scribe v3 Live is also available in ElevenAgents. Existing developer integrations can upgrade by changing the model identifier to scribe_v3 or scribe_v3_live.
A silence timer forces voice agents to choose between answering quickly and risking interruptions, or waiting longer. Scribe v3 Live uses conversation context alongside audio instead. ElevenLabs says this helps avoid cutting off callers halfway through a number. The system already used in ElevenAgents is now available through ElevenAPI.
Speaker labeling, also called diarization, arrives in the live model for the first time. For recordings with three or more speakers, ElevenLabs reports 17% fewer labeling errors with Scribe v3 than Scribe v2.
For live audio, ElevenLabs emphasizes semantic word error rate: it counts an error only when the transcript’s meaning changes. The company favors that metric because streaming providers format numbers, punctuation and capitalization differently.
The multilingual tests used an internal version of FLEURS, a benchmark of short, clean, read speech. ElevenLabs weighted all 102 languages equally and cautions against direct comparisons with scores that pool clips. Speaker-labeling tests used a separate dataset of meetings, calls and voice notes across 16 languages, which the company plans to open-source.
According to ElevenLabs, energy supplier Eneco tested the live model before launch. Its agent recognized Dutch postcodes correctly on the first attempt 40% more often and house numbers 15% more often. The launch post does not give the underlying success rates.
Both models process data in the US by default, with isolated EU, India and Singapore environments for enterprise customers. ElevenLabs also offers private deployment and Zero Retention Mode, which deletes audio and text when a request finishes.
Loading discussion...
Join the conversation
Share a situation where the exact wording would matter.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.