đïž ElevenLabs launches Scribe v3, its most accurate speech-to-text models yet
ElevenLabs released Scribe v3 for recorded audio and Scribe v3 Live for real-time transcription. The company says v3 makes 27% fewer word errors than Scribe v2 across 102 languages on its FLEURS tests, and has the lowest speaker-labeling error of any model it tested, including 17% fewer errors than v2 when three or more people are talking. Scribe v3 Live cuts meaning-changing errors by 43% versus Scribe v2 Realtime, adds live speaker labels for the first time, and uses conversation context for turn detection instead of a simple silence timer. Both models are in the ElevenLabs API, and v3 Live is in ElevenAgents.
â Cleaner transcripts and better "who said what" matter because a missed digit or a wrong speaker can break a voice agent or a meeting summary.
#AI
ElevenLabs released Scribe v3 for recorded audio and Scribe v3 Live for real-time transcription. The company says v3 makes 27% fewer word errors than Scribe v2 across 102 languages on its FLEURS tests, and has the lowest speaker-labeling error of any model it tested, including 17% fewer errors than v2 when three or more people are talking. Scribe v3 Live cuts meaning-changing errors by 43% versus Scribe v2 Realtime, adds live speaker labels for the first time, and uses conversation context for turn detection instead of a simple silence timer. Both models are in the ElevenLabs API, and v3 Live is in ElevenAgents.
â Cleaner transcripts and better "who said what" matter because a missed digit or a wrong speaker can break a voice agent or a meeting summary.
#AI
