Scribe v3 by ElevenLabs

Scribe v3 by ElevenLabs is a speech-to-text model for developers building on calls, meetings, interviews, and media. It reports 27% fewer word errors than Scribe v2 across 90+ languages and is designed to stay accurate amid noise, accents, and crosstalk. The API labels up to 32 speakers, tags audio events and emotion, supports custom vocabulary boosting and entity detection, and handles files up to 15 hours. A low-latency variant, Scribe v3 Live, targets real-time use cases.