Scribe v2 specializes in drug names and clinical dictation

ElevenLabs is extending Scribe v2 with a version specialized for medical transcription. Scribe v2 Medical targets medication names, anatomy, pathology, and clinical dictation while retaining the general model's features and pricing.

Speech recognition specialized for clinical audio

Similar-sounding drug names, dosages, units, and dense anatomical vocabulary expose Speech to Text models to errors that can alter the meaning of a transcript.

Scribe v2 Medical is a version of Scribe v2 fine-tuned specifically for this material. The batch model uses the same API as Scribe v2, including speaker diarization, timestamps, entity detection, and keyterm prompting. It supports more than 90 languages.

ElevenLabs' documentation defines a narrower intended purpose than clinical decision-making: transcripts are designed as drafts for review and correction by healthcare professionals or other authorized users. The model is not intended to interpret medical information, diagnose conditions, or recommend treatment. 4.9% WER on MedDictate

On MedDictate, a clinical dictation benchmark covering English, French, and German, ElevenLabs reports an overall WER of 4.9% for Scribe v2 Medical versus 7.6% for base Scribe v2. GPT Transcribe reaches 5.9% and Muse Voice Transcribe 7.8% in ElevenLabs' evaluation.

Results vary by language. Scribe v2 Medical scores 3.0% in English, 8.3% in French, and 7.2% in German. GPT Transcribe reaches 3.9%, 7.8%, and 9.6%, respectively. OpenAI's model therefore remains slightly ahead in French on this benchmark, while Scribe v2 Medical records the lowest overall WER among the models compared.

On MedTerm, which focuses more specifically on clinical terminology, ElevenLabs reports 77.7% term recall and a 10.4% Term-WER for Scribe v2 Medical. Base Scribe v2 reaches 77.1% and 10.7%, while GPT Transcribe records 74.0% and 12.3%. Isolated terms remain harder

A third benchmark published by Eka Care contains 3,619 English clinical samples spanning drug and condition names, clinical sentences, and conversations between clinicians and patients.

In ElevenLabs' evaluation, Scribe v2 Medical reaches 6.50% SemWER and 6.02% kwWER. Base Scribe v2 records 7.35% and 7.04%, while Deepgram Nova-3 Medical reaches 7.79% and 7.65%.

Medical terms spoken without surrounding context remain more difficult. ElevenLabs reports a 14.3% WER on isolated terms, compared with 7.5% when those terms appear inside full sentences. Keyterm prompting can provide the model with terms likely to appear in the audio and bias transcription toward them. No measured regression on everyday speech

The medical fine-tune does not produce a measurable decline on everyday speech in ElevenLabs' published test. Across 6,000 non-medical Common Voice samples, both Scribe v2 Medical and base Scribe v2 record a rounded WER of 5.3%.

The model is also billed at the same rate as Scribe v2. Developers can select `scribev2medical` through the Speech to Text API. It currently operates through the batch endpoint rather than realtime transcription. Zero retention for protected health information

For workflows involving protected health information, Scribe v2 Medical is HIPAA-eligible for Enterprise customers with a Business Associate Agreement and Zero Retention Mode enabled.

Under this configuration, ElevenLabs says audio input and generated text are deleted immediately after the request completes. Any subsequent retention of the transcript is controlled by the application using the API.