Vibeleaderboard
← All Intel
Intel / article

Elevenagents For Healthcare

Source
elevenlabs.io
Author
ElevenLabs editorial sitemap
Date

Context

The new piece in ElevenLabs' healthcare package for its voice agents is Scribe v3 Live Medical, a speech-to-text model that transcribes callers as they speak and is tuned for drug names and clinical terms (ElevenLabs' announcement). ElevenAgents is the company's platform for building phone and voice agents, and this model is the step that turns what a caller says, such as a member ID or a medication name, into text the can act on. ElevenLabs already sold medical transcription for recorded audio; this brings it to live calls. Teams can also give an agent up to 10,000 terms of their own, such as plan, therapy, device and location names, and the company says the model follows a caller who switches from English to Spanish mid-call.

The accuracy figures come from ElevenLabs' own testing, and the most concrete one is smaller than its headline. On Eka's dataset of more than 3,600 English medical recordings, the company reports a word error rate of 11% for the medical model against 11.54% for Scribe v3 Live, its general real-time model. Word error rate counts transcription mistakes against the number of words actually spoken, so both models make about 11 errors per 100 words, and the advertised "4.7% fewer errors" is a relative cut of roughly half a percentage point. On PriMock57, a set of doctor-patient conversations, ElevenLabs says the model had the lowest error rate among the medical models it tested, with 17% fewer errors than the next closest. The post does not name those models or give their rates, and it notes that the comparison does not include every medical transcription model. No independent test of these results is cited.

Key takeaways · AI-distilled
  • Cadence, which monitors chronic-condition patients, says agent calls after abnormal vital-sign readings cut median time to reach a patient from nearly two hours to 3.5 minutes, and patients answered the agent 47% more often than nurses.
  • In ElevenLabs' internal tests, Scribe v3 Live Medical had 17% fewer errors than the next-best real-time medical model on PriMock57, and 4.7% fewer than the general Scribe v3 Live on Eka's set of more than 3,600 recordings (11% vs 11.54%).
  • Teams can supply up to 10,000 of their own terms, such as plan, therapy and device names, and the transcriber follows callers who switch languages mid-call, for example from English to Spanish.
  • ElevenLabs says that in a study of its agent deployments, a 16% average cut in end-to-end latency correlated with a 26% lift in conversation success rate.
  • HIPAA-eligible deployments are available to Enterprise customers who sign a BAA and turn on Zero Retention Mode, under which audio and transcripts are not stored on ElevenLabs infrastructure.
Terms in this piece · Glossary
  • AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
  • benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
Read the source elevenlabs.io
More from ElevenLabs editorial sitemap
Recommended reads
Comments

Checking sign-in…

Loading comments…