ElevenLabs has released Scribe v3 for recorded audio and Scribe v3 Live for real-time transcription through its API and ElevenAgents platform. Both target difficult conversations with background noise, overlapping speakers and specialized vocabulary across 102 languages.
The company says Scribe v3 reduces word errors by about 25% from its predecessor and improves speaker labeling in larger groups. Developers can preload thousands of terms such as product or medication names, define formatting rules in natural language and handle language changes within a sentence. Enterprise options include EU-isolated servers and a zero-retention mode.
Scribe v3 Live also considers conversational context when deciding whether a speaker is finished. That is intended to prevent voice agents from interrupting people during a thoughtful pause or while they slowly read a number, a task that simple silence timers handle poorly.
Artificial Analysis placed Scribe v3 near the top of its recorded-audio accuracy test, behind only an Alibaba preview model. The same evaluation found it mid-pack on speed and price; Microsoft and xAI alternatives ran much faster with slightly more errors and cost about half as much. The benchmark does not test real-time turn-taking, speaker labeling or rare-language behavior, so buyers must weigh those features separately from raw word-error rates.