Case study

Cut Video Gaps: an Electron desktop app that removes silence and bad takes with offline AI

A macOS app for YouTubers, creators and podcasters that trims dead air, filler words and bad takes from raw recordings, then hands a clean cut to the editor.

  • Electron
  • React
  • TypeScript
  • FFmpeg
  • faster-whisper
  • Python
  • macOS
Cut Video Gaps editor: a 44-second clip with 7 silences and a bad take marked on the timeline, cut down to 29 seconds

The problem

The first edit of a talking-head video or podcast is mostly removal: scrubbing long raw takes for silences, restarts and every "um". It is slow, and it is the same work every time.

Cloud tools solve it by uploading the recording, which is a problem for long files and for anyone who cannot send their footage to a third party.

What we built

  • Silence trimming with FFmpeg, with the threshold set per file from the room's noise floor and an override when you need one
  • Offline AI (faster-whisper plus transcript alignment) that finds filler words and bad takes, including in Hindi and Punjabi
  • Lossless hand-off to Final Cut Pro: FCPXML that points at the original file, frame-accurate with source timecode kept and no re-encode
  • Every proposed cut shown on a timeline, so the editor stays in control

Engineering decisions

  • Frame-accurate re-encode, no stream copy

    Cutting on keyframes is faster but lets audio and video drift out of sync, so video exports always re-encode at the exact frame. Audio-only files are the one exception.

  • faster-whisper over whisper.cpp

    Accuracy in Hindi and Punjabi, hesitation sounds included, was a hard requirement, so the AI runs on faster-whisper in Python, compiled into the app with PyInstaller.

  • Self-hosted Mac build runners

    Cloud macOS runners bill minutes at 10x, and a background service cannot answer keychain prompts. Releases are built on our own Macs with a temporary signing keychain.

Where it stands

  • Live on macOS for Apple Silicon and Intel: signed, notarized and auto-updating
  • In the demo clip on the product site, 44 seconds of raw footage became 29 seconds after one pass
  • Freemium with paid tiers; Windows is planned

Need something similar?

What we learned, written up

More case studies