Skip to content

ElevenLabs Scribe v3 Live Medical gets drug names right on live calls

Comic-style nurse in a headset at a night clinic desk as a glowing sound wave turns into floating pills and pill bottles Models

ElevenLabs has released Scribe v3 Live Medical, a real-time speech-to-text model built for voice agents in healthcare. It launched on November 17, 2026, inside ElevenAgents, the company's platform for building voice agents. The model is tuned to catch medication names, procedures and symptoms that general models often mishear.

ElevenLabs gives an example from its own testing. One speech model heard the diabetes drug sitagliptin as citalopram, an antidepressant. On a live call, an agent could put that wrong drug on a patient's medication list.

What Scribe v3 Live Medical is

The model is a medical version of Scribe v3 Live, the real-time model ElevenLabs released a week earlier. ElevenLabs already offered medical transcription for recorded audio, such as dictation and visit recordings. This is its first medical model that works during a call.

That matters because many healthcare agents work in real time. ElevenLabs lists scheduling lines, medical information lines, eligibility checks and claims follow-ups. These calls can't wait for a recording to finish before the agent acts.

The model is also meant for every call a team handles, including ones with no clinical details. ElevenLabs says it keeps its accuracy on everyday speech, so teams don't need a second model for rescheduling or office-hours questions. Three features target the calls where details matter most.

What Scribe v3 Live Medical can do

The features focus on common problems in patient and provider calls:

  • Custom terms. Teams can load up to 10,000 of their own terms, such as provider names, clinic names or a drug portfolio. The model checks the list only when a caller seems to say one of them. With a list loaded, it made up to 48% fewer errors on medical terms in ElevenLabs' tests.
  • Turn detection. Callers often read a member ID or date of birth in pieces. The model recognizes when a number isn't finished, so the agent waits instead of cutting in.
  • Language switching. It supports more than 90 languages and follows callers who switch mid-call, for example to describe a symptom in their first language.

Transcription takes about 150 milliseconds, the same as Scribe v3 Live, so the medical tuning adds no noticeable delay. ElevenLabs backs the accuracy claims with a long set of benchmarks.

How Scribe v3 Live Medical scores on benchmarks

ElevenLabs tested the model in real-time mode against medical and general-purpose models from other providers. The main test was PriMock57, 57 unscripted mock doctor visits of about nine minutes each over phone audio with background noise. The second was Fareez OSCE, 272 recorded patient interviews from a different team in another country.

ElevenLabs benchmarks: Scribe v3 Live Medical has 12.8% WER on PriMock57 vs 15.5% for GPT-Live-Transcribe, and 0.6% semantic WER on Fareez OSCE
ElevenLabs internal testing, November 2026

On PriMock57, Scribe v3 Live Medical had a word error rate of 12.8%. The next medical model, AssemblyAI's Universal-3.6 Pro Realtime in medical mode, scored 15.5%, as did OpenAI's GPT-Live-Transcribe. That is about 17% fewer errors. Google's Gemini 3.5 Transcribe Live came last at 18.7%.

Fareez OSCE uses semantic word error rate, which counts only mistakes that change the meaning. Scribe v3 Live Medical scored 0.6%, the lowest among medical models. Microsoft's general-purpose MAI-Transcribe-2 Streaming did slightly better, at 0.5%.

ElevenLabs also compared the model with Scribe v3 Live. On the Eka Medical ASR set of drug names and medical terms, it scored 11% against 11.54%. On everyday English sentences from Common Voice and FLEURS, the two models stayed within a few tenths of a point. On Corti's MED-Term set, it scored 2.24% in English, 2.81% in German and 3.24% in French. Early customers say the gains show up on real calls.

What healthcare teams say about Scribe v3 Live Medical

ElevenLabs quoted two early users. Pfizer's Idin Jafari said the agent is "more accurate on the brand and generic names" in Pfizer's calls. Pfizer uses it together with Eleven v4 Turbo, ElevenLabs' voice model.

Vintage Labs AI makes pre- and post-procedure calls for large academic health systems. Co-founder Shantanu Ghorai said turn detection now waits for patients to finish their date of birth. He added that nurses can now trust the call summary without replaying the recording.

ElevenLabs puts the numbers in call terms. A 10-minute intake call runs about 1,500 words. At the PriMock57 rates, Scribe v3 Live Medical would get about 40 more words right per call than the next medical model. Patient data on those calls brings its own rules.

Scribe v3 Live Medical availability and privacy

The model is available now in ElevenAgents, under Settings > Automatic Speech Recognition. It is HIPAA-eligible for enterprise customers with a business associate agreement and Zero Retention Mode turned on.

ElevenAgents is covered by SOC 2 Type II, ISO 27001 and PCI DSS Level 1 programs. Patient data is stored in the US by default, with isolated environments in the EU, India and Singapore for enterprise teams. ElevenLabs has not listed a separate price for the medical model.

Is Scribe v3 Live Medical the best medical transcription model

On ElevenLabs' numbers, it is the most accurate real-time medical model the company tested. Its lead on PriMock57 is clear, and it does not give up accuracy on everyday speech. The 10,000-term list and turn detection address real problems in patient calls.

The caveats are familiar. All results come from ElevenLabs' own tests of a chosen set of rivals, and on Fareez OSCE a general Microsoft model came out slightly ahead. For teams already building health agents in ElevenAgents, switching is a single setting, so the model is worth testing on their own calls.

Sources

Portrait of Felix Laurent, author at thehype.news

Felix Laurent

News Writer at thehype. Felix is a journalist with experience in marketing. He covers artificial intelligence and emerging technologies, focusing on the ideas, innovations, and trends shaping the future of tech.

More articles by Felix Laurent →
On air · radio.thehype.news ↗ AI news radio — 24/7