ElevenLabs releases Scribe v3

ElevenLabs' post on 10 November 2026 launches Scribe v3 for recorded audio and Scribe v3 Live for realtime audio. ElevenLabs says Scribe v3 makes 27% fewer word errors than Scribe v2 across 102 languages on its internal FLEURS test, and that it had the most accurate speaker labels of any model they tested. On recordings with three or more speakers, ElevenLabs says Scribe v3 makes 17% fewer diarization errors than Scribe v2, and 12% fewer overall.

Scribe v3 Live includes speaker diarization for the first time, and turn detection that uses the conversation as well as silence. ElevenLabs says that, across 102 languages, Scribe v3 Live reduces semantic word error rate by 43% and word error rate by 31% compared with Scribe v2 Realtime. It also says Live makes 39% fewer diarization errors than the next-best realtime model they tested. On the English portion of a 16-language set of real meetings, calls, and voice notes, ElevenLabs reports a Live diarization error rate of 3.1%, against 5.9% to 7.3% for the other realtime models in that test. Those figures are ElevenLabs'.

Scribe v3 can take up to 1,000 key terms, such as names and jargon, and can return numbers as digits or words. Scribe v3 Live adds a multilingual mode for a conversation that switches language mid-sentence. Scribe v3 Live is in ElevenAgents. Both models are the default in the ElevenLabs app and are in ElevenAPI. Existing integrations change the model id to scribe_v3 or scribe_v3_live. ElevenLabs says both run in the US by default, with isolated EU, India, and Singapore environments for enterprise customers, and that Zero Retention Mode deletes audio and text when a request completes.

Who it's for: someone transcribing a recording or a live call who already uses ElevenLabs, and who needs speaker labels or a realtime transcript.

Fits: Best for AI Voice stays ElevenLabs, for a voiceover that has to sound like a performance. Scribe v3 transcribes speech. It does not change that voiceover. Best for Podcast Editing stays Descript, which edits by the transcript. A transcript model is not that editor. The voice guide and the podcast guide are unchanged. The verdict is unchanged.