Medical transcription that never has to leave the room.
VAYS MED brings VAYS's on-device transcription to clinical work: turn a consultation into a clean transcript with the speakers separated, without sending patient audio to the cloud.
VAYS MED brings VAYS's on-device transcription to clinical work: turn a consultation into a clean transcript with the speakers separated, without sending patient audio to the cloud.
Turn a consultation into a clean, structured transcript, on the device, in Polish, English and more.
Speaker separation labels the doctor and the patient, so the record reads as a conversation instead of a wall of text.
Audio and transcripts stay on your hardware, built for environments where patient data must never leave the room.
Accuracy is easy to claim and hard to check, so here is the measurement. ADMEDVOICE is a public corpus of Polish medical speech published in Nature Scientific Data (CC BY 4.0): radiology reports, oncology notes, referrals, resuscitation records. We ran the exact on-device engine VAYS ships against it. To our knowledge this is the first published measurement of this engine on Polish medical speech.
Across 9,797 recordings and 16.3 hours of audio. Character error rate 9.4%.
On an Apple Silicon Mac, fully offline. A 20 minute consultation transcribes in well under a minute.
The whole benchmark ran on the device with no network path, on the same code a clinic would run.
| Specialty | Word error rate |
|---|---|
| Referrals | 10.4% |
| Recommendations | 11.6% |
| Head and neck radiology | 11.4% |
| Oncology | 19.2% |
| Thyroid radiology | 23.3% |
| Resuscitation | 25.8% |
| Cardiology | 44.0% |
Difficulty tracks vocabulary density. Referrals and recommendations are short and formulaic. Cardiology is the hardest case in the corpus, with dosages, parameters and units in a single breath. Part of that tail is the corpus itself: 80 recordings have mismatched audio and reference text, and with those filtered out cardiology lands near 37%.
Read this for what it is. These are speech engine numbers, not a claim about VAYS MED as a product. ADMEDVOICE is single-speaker dictation, while a consultation is two people talking, so the task is genuinely different. VAYS MED is in development and has not yet been measured on real consultations. When it is, that number goes on this page too, whatever it turns out to be.
VAYS MED runs on the same private-AI foundation as VAYS: fully on-device, with privacy and control on your side. No backend and no third-party cloud in the default path.
See VAYS KIT →VAYS MED is in development and is not a medical device. It's a documentation and transcription aid; clinicians remain responsible for records. Processing runs on-device and you stay in control of your data, which matters under GDPR, including special-category health data.
We're talking to clinicians and clinics helping shape it. Reach out for early access.