Extract Clinical Notes from Audio
From exam-room audio to a structured clinical note.
Builders: Senthil P.
Open-Source AI ยท Runner-upHealth accessGemmaOpen-Source AI
A reproducible pipeline from consultation audio to speaker-labeled transcript to a structured clinical note, built around evaluation.
Replay demo
Demo only, not a clinical tool. Simulated consultation with actors from the PriMock57 dataset. No real patient data.
Builder: Senthil Prabakaran
Full consultation (09:04)Synced transcript
Loading transcript and note.
Pipeline facts
- Whisper base int8 on CPU, no diarization.
- WER 27.2% raw, 26.9% corrected.
- Gemma 4 31B at temperature 0, strict JSON schema.
- 33 facts, 100% evidence grounding, 100% recall of clinician-highlighted facts.
Limitations
- Simulated, English only, one encounter.
- Base Whisper, no diarization.
- Draft for clinician review, not a medical device and not medical advice.
How it works
Gemma 4 31B; ships with two Kaggle datasets pairing encounter audio with clinical notes.
What we would test next
Next step: end-to-end note generation with clinical review.
From the judges' pre-demo code reviews.
Something off, or shipped an update since Saturday? Tell us