ElevenLabs launches Scribe v2 Medical, a clinical speech-to-text model

ElevenLabs made Scribe v2 Medical generally available on September 22, 2026, a speech-to-text model fine-tuned specifically for clinical audio. It cuts word-error rate on medical dictation by roughly 35% compared to the company's base Scribe v2 model while matching that model's accuracy on ordinary speech.

What's new

Scribe v2 Medical is a fine-tune, not a from-scratch model. ElevenLabs built it to handle the specific failure points of clinical transcription: drug names that differ by a single syllable (the company cites hydroxyzine and hydralazine as an example pair), dense strings of dosages and units, and anatomy and pathology terms arriving in rapid sequence. ElevenLabs frames the stakes plainly: a transcription error in this setting can end up in a patient's chart.

The model is available now on the Speech to Text API under the model ID scribe_v2_medical, running on the batch endpoint rather than the realtime one, so it fits dictation and post-visit transcription workflows more than live conversation capture. It carries the same pricing, feature set, and API surface as base Scribe v2, and any existing Scribe v2 integration can switch by changing the model ID alone. ElevenLabs reports the accuracy gains hold across English, French, and German, tested against its internal MedDictate and MedTerm benchmarks.

For healthcare customers, the model is HIPAA-eligible for Enterprise accounts that have a Business Associate Agreement in place and Zero Retention Mode turned on. With that mode active, both the audio input and the text output are deleted immediately once a request completes, which matters for any covered entity that can't have patient audio persist on a vendor's servers.

Context

Speech-to-text has become a crowded, and increasingly specialized, corner of the AI market. General-purpose transcription models from OpenAI, Google, and others handle everyday speech well, but medical dictation is a narrower, higher-value niche already served by dedicated vendors like Nuance (now part of Microsoft) and Suki. ElevenLabs entered speech-to-text more recently than it entered voice generation, and Scribe v2 Medical is its first model built for a single vertical rather than for the general population of dictation and transcription users.

Why it matters

The 35% WER reduction, if it holds up in outside testing, is the kind of number that decides whether a hospital or clinic actually trusts an AI transcription tool in the exam room instead of just piloting it. General-purpose speech models tend to fail in predictable ways on clinical audio: they mishear drug names, guess wrong on dosage units, and stumble on pathology terminology that rarely shows up in their training data. A model tuned specifically against those failure modes, priced the same as the general model and swappable with a one-line model-ID change, lowers the switching cost for any developer already on ElevenLabs' API.

The HIPAA-eligibility and Zero Retention Mode details are the other half of the pitch. Clinical customers care as much about where the audio goes as they do about accuracy, and ElevenLabs is positioning this release to satisfy both requirements at once rather than treating compliance as a separate enterprise add-on.