From the source
ElevenLabs is releasing Scribe v3 Live Medical, a real-time speech-to-text model built on Scribe v3 Live and tuned for clinical language, available now in ElevenAgents.
- The company says it already offers medical transcription for recorded audio, but healthcare teams also run scheduling, information lines, eligibility checks and claims follow-ups in real time, and those calls cannot wait for a recording to end.
- Medication names and doses rarely appear in everyday speech, so a general real-time model is more likely to get them wrong; ElevenLabs cites a test where a speech model transcribed sitagliptin as citalopram.
- On PriMock57, a public set of 57 mock live primary care consultations, Scribe v3 Live Medical had the lowest word error rate of the medical models tested at 12.8%, which ElevenLabs says is 17% fewer errors than the next closest.
- On Fareez OSCE, 272 recorded patient interviews, it had the lowest semantic WER of the medical models tested at 0.6%.
- Teams can load up to 10,000 custom terms, and with a list loaded ElevenLabs measured up to 48% fewer errors on medical terms; the model supports more than 90 languages and follows mid-call language switches.
- ElevenLabs says the results come from its internal testing of selected providers in real-time mode, with each model in its best available configuration, and do not include every available model.






