Skip to main content
Back to timeline
ElevenLabs BlogSource publication:

ElevenLabs moves medical transcription into live calls: Scribe v3 Live Medical posts 12.8% WER on PriMock57, 17% fewer errors than the next medical model

Synopsis

ElevenLabs released Scribe v3 Live Medical, bringing real-time speech transcription tuned for clinical language to call agents in ElevenAgents. Its earlier medical transcription covered recordings only, while drug names and doses in live calls are what general models most often mishear. On PriMock57, 57 mock primary care consultations, the model posted 12.8% word error rate, 17% fewer errors than the next closest medical model tested, and up to 48% fewer errors on medical terms with a custom term list loaded.

Source-provided article image: Scribe v3 Live Medical helps conversational agents get clinical details right
ElevenLabs

Interpretation

Voice agents on live calls can now call a transcription model optimized for clinical language, writing the drug names, doses and symptoms a patient says into the call record for clinicians to review. ElevenLabs previously offered medical transcription for recorded audio such as dictation and visit recordings, but scheduling, medical information lines, phone eligibility checks and claims follow-ups run in real time and cannot wait for a recording to end. On PriMock57, a public set of 57 mock primary care consultations, Scribe v3 Live Medical had the lowest word error rate of the medical models tested at 12.8%, 17% fewer errors than the next closest one.

The model is more accurate on clinical details while holding its accuracy on everyday speech, so teams can use it for every call instead of switching models by call type. Medication names and doses rarely appear in everyday speech, which makes them the words a general real-time model is more likely to get wrong; the material notes a speech model transcribing sitagliptin as the antidepressant citalopram in testing. On Eka Medical ASR, Scribe v3 Live Medical reached 11% semantic word error rate versus 11.54% for Scribe v3 Live, and the two perform about the same on the English sets of Common Voice and FLEURS.

The model recognizes when a caller has not finished a number, waiting for the rest of a member ID or date of birth before replying, and follows language switches mid-call. Callers often read a member ID or date of birth in pieces, pausing between groups of digits, and an agent that treats every pause as the end of a turn jumps in mid-number and makes the caller start again. The model supports more than 90 languages and follows switches during a call; on Corti's MED-Term benchmark its semantic word error rate was 2.24% in English, 2.81% in German and 3.24% in French.

Perspective

The demonstrated scope is ElevenLabs' internal testing of selected providers' real-time models, covering the 57 mock primary care consultations of PriMock57, 272 recorded patient interviews in Fareez OSCE, Eka Medical ASR, the English sets of Common Voice and FLEURS, and MED-Term in English, French and German; results are in English unless noted. The model is available in ElevenAgents, and is HIPAA-eligible for enterprise customers with a business associate agreement and Zero Retention Mode enabled.

The material states that these results come from ElevenLabs' internal testing of selected providers in real-time mode, with each model in its best available configuration, and do not include every available model; PriMock57 and Fareez OSCE are mock or recorded settings rather than live production calls.