Open models transcribe 8h38m of mock GP consultations in 18m38s on a single GPU — 27.8x faster

MaziyarPanahi · x · 2026-11-19

Maziyar Panahi demoed transcribing and drafting notes for 57 mock GP consultations (PriMock57 dataset, 8h38m of audio) in 18m38s on a single GPU using only open models — 27.8x faster than real time, with each draft ready 20s after its call.

He clarified the calls are mock consultations with clinicians playing patients (CC BY 4.0), not real patients, and every draft is for doctor review and sign-off. Many doctors still write notes late into the night, which this workflow aims to relieve.

Related event: Open-source models transcribe 8.7 hours of clinic audio in 18 minutes on one GPU(2 posts)→

Original post →

More from Apps

Apps channel →