Ahmad ZushanAI · Systems · Platform

Home/2026 builds/№ 37

37 · Voice & language

Scribe Studio — local speech-to-text and OCR workbench

A FastAPI gateway supervises two worker processes, speech and OCR, inside a Windows job object, leases the GPU between them and unloads the idle engine when VRAM runs short; the interface opens as an Edge app window from a tray icon.

PeriodSep 2026StatusLocal appAreaVoice & language

Highlights

  1. Whisper large-v3-turbo at 9.9× real time in int8_float16 on a GTX 1650, with a fix for batched inference dropping speech after the last timestamp token.
  2. Multi-script OCR: a Hindi pipeline detects the lines, then Urdu and English recognisers re-read each crop and the highest confidence wins (96–98% on test pages).
  3. Groundwork for Nastaliq: a synthetic corpus and rendered line images, because stock Urdu OCR fails on Nastaliq with a page error rate above 50%.