← Work

Solo build · 2026 · Full-stack, solo

Whisper Scribe

Local-only speech-to-text — your audio never leaves the machine.

Source on GitHub
What it is
A private transcription tool: import a recording, get a clean transcript with click-to-jump timestamps, inline editing, and five export formats — fully offline after a one-time model download.
The problem
I wanted a Whisper-Flow-style transcriber, but with a hard privacy line: no cloud call, no vendor lock-in, fully offline once the model weights are downloaded.
What I built
A FastAPI backend running faster-whisper CPU inference with local speaker diarization, a dependency-light vanilla-JS single-page frontend, and PyAV for audio/video decode. Verified end-to-end with real inference — and real cross-platform bugs fixed — on both macOS and Windows.
Stack
PythonFastAPIfaster-whisper (CTranslate2)Vanilla JSPyAVsherpa-onnx
Results
  • Tested end-to-end — real inference, API and browser — on macOS and Windows
  • Speaker diarization, fully local
  • Five export formats, click-to-jump editing
What it shows
Taking a product spec to a genuinely working full-stack tool solo — then debugging it for real conditions on hardware it wasn't written for.
← Back to all work