Dictate the note, get the note. Built for clinicians speaking findings, prescriptions and case summaries, with every drug name, dose and term intact.
Upload a dictation · or hold Space and dictate
Indian clinicians dictating drug names, findings and case notes. Same audio to both models; listen, then read. Underlined: the term that differs.
Audio from the Eka Medical ASR Evaluation Dataset (ekacare, MIT licence), English test split. General-purpose model: Whisper large-v3, default English transcription. Outputs shown verbatim.
Built for the dictating clinician: the note comes out the way it was spoken.
A speech model post-trained on clinicians speaking medicine, so anatomy, pathology, procedures and pharmacology come out as the words that were dictated, not their nearest everyday sound-alike.
Terminology survives the transcriptGeneric and brand names, milligram and microgram doses, frequencies and routes are the parts of a dictated note that matter most and that general-purpose recognisers break first. They are treated as first-class vocabulary here.
“metformin 500 mg twice daily” stays exactly thatAccuracy is reported on held-out recordings of Indian clinicians the model never saw in training, scored the same way as the clips above. No demo-only tuning.
Word error rate down by a quarter after post-trainingA compact model on a single GPU turns a 30-second dictation into text in well under a second of processing.
Server RTF below 0.1 on live clipsSentence casing, punctuation and section structure are restored on the server so the dictation drops straight into an EMR field, prescription or discharge summary without a clean-up pass.
Copy, paste, signAudio is processed on infrastructure we run, never sent to a third-party speech API. Deployable inside a hospital network or a dedicated cloud tenancy, with no model or vendor names exposed to end users.
Your clinicians’ voices stay with you