ElevenLabs launches ElevenAgents for Healthcare with Scribe v3 Live Medical: 17% fewer word errors than the next medical model on PriMock57 doctor-patient calls
Synopsis
ElevenLabs released ElevenAgents for Healthcare with Scribe v3 Live Medical, its first real-time medical transcription for agents, so an agent can recognize drug names and clinical terms during a live call. Previously ElevenLabs offered medical transcription only for recorded audio, leaving live-call recognition and turn-taking unaddressed. On the PriMock57 doctor-patient benchmark it had the lowest word error rate among the medical models ElevenLabs tested internally, 17% fewer errors than the next closest; on Eka's dataset of more than 3,600 English medical recordings it made 4.7% fewer errors than the general real-time model Scribe v3 Live (11% vs. 11.54%).
Interpretation
An agent on a healthcare line can now transcribe clinical language as the call happens, so drug names, complex terms and member IDs need not be repeated. Before this, ElevenLabs' medical transcription covered recorded audio only, with no counterpart for live calls. On the PriMock57 doctor-patient benchmark, Scribe v3 Live Medical had the lowest word error rate among the medical models ElevenLabs tested internally, 17% fewer errors than the next closest.
The transcription is built on the general real-time model Scribe v3 Live and tuned for drug names and clinical terms. Scribe v3 Live is the general real-time base version that the medical variant adjusts. On Eka's dataset of more than 3,600 English medical recordings, Scribe v3 Live Medical made 11% errors against 11.54% for Scribe v3 Live, 4.7% fewer.
Perspective
The demonstrated scope is real-time call transcription and agent behavior control in healthcare settings, including up to 10,000 custom terms, following a caller who gives a member ID in English and then asks a question in Spanish, turn-taking detection and transfer to a person; it is aimed at teams deploying healthcare call agents.
The word error results come from ElevenLabs' internal testing on a selection of real-time models, not every medical transcription model available; the reported 16% average latency reduction correlated with a 26% lift in conversation success rate comes from a study of ElevenAgents deployments and is a correlation.
