All news

ElevenLabs releases Scribe v3 and Scribe v3 Live for recorded and streaming transcription

Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .

Voice agents and meeting transcription will become more accurate in multilingual conversations, while pauses and speaker changes will be recognized more naturally.

In ElevenLabs' internal tests across 102 languages, Scribe v3 reduces word errors by 27% compared with v2, while Scribe v3 Live reduces errors by 43% relative to v2 Realtime. The streaming version gains speaker labels for the first time and uses conversational context to detect the end of a turn, avoiding interruptions during long pauses. For recorded audio, diarization errors fell by 12% overall and by 17% with three or more speakers. The models support key-term lists, formatting rules written in natural language and language switching within a sentence. Scribe v3 Live is already available in ElevenAgents, and both models are available through the API; some results were measured on an internal ElevenLabs dataset that the company promises to open-source.