model releaseElevenLabs

ElevenLabs releases Scribe v3 Live Medical, real-time clinical transcription with 12.8% WER on PriMock57

TL;DR

ElevenLabs has released Scribe v3 Live Medical, a real-time speech-to-text model tuned for clinical language and available now in ElevenAgents. The company claims it posted the lowest WER of the medical models it tested on PriMock57 (12.8%).

4 min read
0

ElevenLabs has released Scribe v3 Live Medical, a real-time speech-to-text (STT) model optimized for clinical language, and it is available now in ElevenAgents. The company says it is its first medical transcription model built for agents handling live calls.

Scribe v3 Live Medical is built on Scribe v3 Live, ElevenLabs' newest real-time STT model. It carries the clinical focus of Scribe v2 Medical, which ElevenLabs already offers for recorded audio such as dictation and visit recordings, into live conversations. Pricing is not yet disclosed.

Features

  • Custom terms: Teams can load up to 10,000 of their own terms, such as provider and clinic names or a pharmaceutical portfolio by brand and generic name. The model draws on the list only when a caller seems to say one of the terms. ElevenLabs claims up to 48% fewer errors on medical terms with a custom list loaded.
  • Number handling: The model recognizes when a spoken number, such as a member ID or date of birth, is not finished, so the agent waits instead of replying mid-number.
  • Language switching: It supports more than 90 languages and follows mid-call switches, so speech in a second language is transcribed instead of dropped.
  • Latency: Scribe v3 Live transcribes speech in about 150 milliseconds. ElevenLabs says that in testing, Scribe v3 Live Medical matches that latency.

ElevenLabs illustrates the problem with an example from its own testing: a speech model transcribed the diabetes drug sitagliptin as citalopram, an antidepressant.

Benchmark results

All results come from ElevenLabs' internal testing of selected providers in real-time mode, with each model in its best available configuration. The company says the comparison does not include every available model. Results are in English unless noted. WER is word error rate. Semantic WER counts only errors that change meaning, so it scores lower than WER.

Benchmark Metric Scribe v3 Live Medical
PriMock57 (57 mock primary care consultations, about 9 minutes each) WER 12.8%
Fareez OSCE (272 patient interviews) Semantic WER 0.6%
Eka Medical ASR (more than 3,600 recordings) Semantic WER 11%
MED-Term, English Semantic WER 2.24%
MED-Term, German Semantic WER 2.81%
MED-Term, French Semantic WER 3.24%

ElevenLabs says Scribe v3 Live Medical had the lowest WER of the medical models it tested on PriMock57, with 17% fewer errors than the next closest one. It also reports the lowest semantic WER of the medical models tested on Fareez OSCE.

On Eka Medical ASR, which focuses on medical terms and branded drug names, ElevenLabs reports 4.7% fewer errors than Scribe v3 Live (11% vs. 11.54% semantic WER). On everyday speech, scored on the English sets of Common Voice and FLEURS, the company says Scribe v3 Live Medical performs about the same as Scribe v3 Live. MED-Term is a public benchmark from Corti.

Customer comments

Idin Jafari, Director, Digital Strategy & Innovation at Pfizer, said: "The interaction experience with Eleven v4 Turbo and Scribe v3 Live Medical is phenomenal. The agent sounds realistic, and it's more accurate on the brand and generic names that come up in our conversations."

Shantanu Ghorai, speaking about Vintage Labs' VIA platform, which makes pre- and post-procedure calls for large academic health systems, said that with Scribe v3 Live Medical it "understands patients better: their medications, their procedures, the names of their doctors."

What this means

A voice agent can only act on what it transcribes, so a misheard drug name can end up in a medication list a clinician reviews. Scribe v3 Live Medical targets that failure point in real-time calls, where the agent cannot wait for a finished recording.

ElevenLabs says the model holds its accuracy on everyday speech, which means one model can serve both clinical and routine calls such as rescheduling. The custom term list and number-completion handling address two specific problems in healthcare calls: rare proper names and IDs read out in pieces. The benchmark figures are ElevenLabs' own, run on a selected set of providers, so teams evaluating the model should test it on their own call audio and terminology.

Related Articles

model release

RadixArk releases DaisyRose-27B, lifting Terminal-Bench 4.1 pass@1 from 5.6% to 12.0%

RadixArk released DaisyRose-27B, its first model, built on Qwen3.8-27B and trained with reinforcement learning on 1,200 curated agent tasks. According to RadixArk, it raises Terminal Bench 4.1 pass@1 from 5.6% to 12.0% and scores higher than the base model on all seven reported benchmarks. Weights ship under Apache 2.0 in BF16, FP8 and NVFP4.

model release

Google releases Genie 4 to AI Ultra subscribers with real-time prompt editing and native audio

Google is rolling out Genie 4, which it calls its most realistic and interactive world model yet, to eligible Google AI Ultra subscribers in Project Genie. The update adds real-time native audio, mid-session text prompting, a world-building agent and a community gallery. Genie 3 also gains sessions of up to 3 minutes.

model release

Microsoft releases Aion 1.0 Instruct, a 2.3B on-device SLM with 32K context under MIT license

Microsoft's Windows Applied Sciences Group has published Aion 1.0 Instruct, a roughly 2.29B-parameter instruction-tuned small language model with a 32,768-token context window under the MIT license. A further optimized version is slated to become the Windows OS SLM by the end of January 2027.

model release

Bria releases RMBG-3.0, a 4-step background remover built on Fibo 2.0, with open weights for non-commercial use

Bria AI has released RMBG-3.0, the third generation of its background removal models and the first built on a generative foundation. It is Fibo 2.0, fine-tuned and distilled to produce an alpha matte in 4 steps. Bria claims it ranks first among six systems in its own benchmark of 234 hard images, with an average rank of 1.51.

Comments

Loading...