ElevenLabs ships Scribe v3 with live speaker diarization
The new speech-to-text models cut word errors by up to 43% over their predecessors and bring real-time diarization and mid-sentence language switching to live audio.
ElevenLabs released Scribe v3 and Scribe v3 Live on 10 November, its most accurate speech-to-text models yet across 102 languages. Scribe v3 handles recorded audio with 27% fewer word errors than Scribe v2. Scribe v3 Live, the low-latency variant for streaming audio, makes 43% fewer semantic errors than Scribe v2 Realtime.
Both models get better at telling speakers apart: 12% fewer diarization errors overall for Scribe v3, rising to 17% on recordings with three or more speakers. Scribe v3 Live brings diarization to real-time transcription for the first time, with 39% fewer diarization errors than the next-best realtime model ElevenLabs tested against. It also adds a multilingual mode for conversations that switch languages mid-sentence.
model_id: "scribe_v3_live"
Scribe v3 Live is live now in ElevenAgents, including a healthcare-tuned version; both models are reachable through the API by switching the model ID. Deployment is available in US, EU, India and Singapore environments.
Why it matters: real-time diarization has been the weak point in live transcription, call centres and meeting tools have had to bolt it on or wait for a recording to finish. Having it in the streaming model itself, plus a model ID swap rather than an integration rewrite, lowers the cost of adding or upgrading transcription in a voice product.
The caveat: the error-rate figures are ElevenLabs’ own benchmarks, not an independently run comparison, and no pricing was given alongside the release.