ElevenLabs has released Scribe v3 Live Medical, a real-time speech recognition model built for healthcare voice agents that handle live calls with patients, health plan members and providers. It is optimized for clinical language such as drug names, doses and procedures, and it is available in early access inside ElevenAgents, the company’s voice agent platform. The goal is simple: an AI agent can only act on what a caller says if it hears the medication name correctly.
Quick Take
- The launch: Scribe v3 Live Medical, announced November 17, built on the new Scribe v3 Live real-time model.
- Use cases: Intake and scheduling calls, claims and prior authorization calls, medical information lines and follow-ups.
- Custom vocabulary: Up to 10,000 custom terms, such as provider, clinic and drug names. With a term list loaded, ElevenLabs says the model made up to 48% fewer errors on medical terms in its tests.
- Languages: More than 90, including mid-call language switches.
- Limits: Early access, available only in ElevenAgents and not through the Speech to Text API. Transcripts are drafts meant for review by a healthcare professional.
In plain English
Phone agents in healthcare listen to callers and then do things: update a medication list, check eligibility or file a claim. If the speech recognition mishears “sitagliptin” as a different drug, the agent acts on the wrong information. ElevenLabs trained a version of its live transcription model specifically on medical language so these agents make fewer of those mistakes.
What happened
In its announcement, ElevenLabs gave a concrete example: in its testing, a speech model transcribed the diabetes drug sitagliptin as citalopram, an antidepressant. Scribe v3 Live Medical carries the clinical focus of the company’s existing Scribe v2 Medical, which handles recorded audio like dictation, into real-time conversations. On PriMock57, a public set of mock doctor visits, ElevenLabs says the new model made fewer errors than any other medical model it evaluated.
The model also uses turn prediction: when a caller reads out a member ID or date of birth in chunks, it recognizes that the number is not finished and waits, instead of interrupting. ElevenLabs says the model is HIPAA-eligible for enterprise customers with a business associate agreement and Zero Retention Mode, while its documentation asks early access users not to send protected health information until HIPAA eligibility is confirmed for this model.
Background
Voice agents are spreading fast across healthcare administration, from appointment booking to insurance follow-ups. Just yesterday we covered August AI’s health agent that calls doctors and fights insurance denials, and AI scribes are winning large contracts, as in Abridge’s VA deal.
Why it matters
In clinical settings, transcription errors are not a cosmetic problem. A wrong drug name can flow into a medication list or a prior authorization. Specialized medical speech models with custom vocabularies are a practical safety layer for the wave of healthcare voice agents now being deployed.
What’s still open
- Independent accuracy results beyond ElevenLabs’ own testing.
- When HIPAA eligibility is confirmed for the early access model and when EU data residency becomes available.
- Pricing for the medical model.
About the analysis
This independent i10x analysis is based on ElevenLabs’ blog announcement and its model documentation. Accuracy figures are company test results and were not independently verified.
Sources (plain text, no links):
- ElevenLabs blog: “Scribe v3 Live Medical helps conversational agents get clinical details right” (November 17, 2026)
- ElevenLabs documentation: Models overview, Scribe v3 Live Medical
i10x Perspective
Vertical versions of general AI models are becoming the norm in healthcare, because general accuracy is not enough when rare words carry the most risk. The careful framing here, drafts for human review and no diagnosis, is also a sign of how vendors are positioning around regulation. More on the Health & Medical Research page.



