Drop a video in, the silences come out.
Export an MP4, or hand the cut timeline to DaVinci Resolve and Premiere.

---

Cutting the pauses out of a talking head video is mechanical work: find the gap, drag the edges in, ripple everything left, do it two hundred more times. autocut is a desktop app that does it in one pass. The footage stays on your disk, there is nothing to install beside it (no Python, no ffmpeg), and what comes out is either a finished MP4 or a timeline your NLE opens without a relink dialog.

brew install --cask cobanov/tap/autocut

• The sliders are instant. The speech model reads the track once, so tuning reshapes a result autocut already holds instead of re-running anything.
• The preview is the export. The player skips removed regions as it plays.
• Multi-track shoots stay in sync. Other angles and a separate sound recorder are cut at the same points, aligned by their embedded timecode.
• FCPXML that relinks itself. Source timecode is preserved, so DaVinci and Premiere bind each clip to its original media.
• Nothing is uploaded. Silero V5 runs locally, ffmpeg ships inside the app.

Install

macOS, Apple Silicon. Homebrew is the easier path, because it clears the Gatekeeper quarantine flag for you:

brew install --cask cobanov/tap/autocut

Installing the .dmg from the latest release by hand works too: drag autocut into Applications, then run xattr -cr /Applications/autocut.app once, or macOS calls the app damaged and refuses to open it. The bundle is not notarized yet, and that command is the whole difference between the two paths.

Windows, x86_64. autocut_X.Y.Z_x64-setup.exe (NSIS) or the .msi, from the latest release. Unsigned, so SmartScreen wants More info, then Run anyway on first launch.

Linux. Build from source for now.

Use

1. Drop a video on the window. MP4, MOV, MKV, WebM and AVI all work.
2. Hit detect silences. Spoken regions turn green, silent ones red.
3. Watch the preview. Space plays and pauses, and what you hear is what exports.
4. Refine. Drag the green edge handles, type exact times in the cuts panel, or click the x on a row to park a keep without deleting it.
5. Export an MP4 to send, or an FCPXML to keep editing.

Slider	Default	What it does
threshold	0.50	How sure the model must be that a chunk is speech. Higher cuts more
pad	0.30s	Breathing room kept on both sides of every spoken region
min silence	100ms	Shorter gaps are not worth cutting, so they stay in
min speech	150ms	Shorter bursts are dropped, which kills clicks and lip noise

Hold shift for fine steps. Scroll the timeline to pan, drag in the navigator below it to zoom. The first detect on a long video takes a moment, because that is the model reading the whole track; every re-detect after it is free.

Multi-track shoots

Two cameras and a recorder on the table are one performance, and they have to be cut identically or they drift apart. Drop your main camera first: that is the reference, the thing you preview and everything else lines up against.

In the export panel, + add under linked tracks takes the other angles and audio files. Files carrying timecode are aligned from it (tc); files without it are assumed to have started together (≈), so type the real offset if they did not. If you recorded clean sound separately, point listen to in the detection panel at it, and the analysis reads the good microphone while the cuts still apply everywhere.

Export FCPXML and the main camera lands on V1, the other video above it, the audio below, every track cut at the same frames. MP4 export still writes one file from the reference clip: a flat video has one picture and one mix, and choosing which angle is on top is an edit, not a cut.

Development

Svelte 5 and TypeScript in front, Rust and Tauri 2 behind, with ffmpeg and ffprobe as bundled sidecars.

pnpm install
pnpm tauri:dev                                # the first build fetches ~200MB of ffmpeg

pnpm test                                     # vitest, the cut algebra
cd src-tauri && AUTOCUT_STUB_SIDECARS=1 cargo test --lib

AUTOCUT_STUB_SIDECARS=1 drops empty placeholders where the ffmpeg binaries go, so a fresh checkout runs the unit tests without fetching them first. Never set it for a real build: the bundle cannot process video, and it says so at compile time.

The version lives in package.json. tauri.conf.json reads it from there, and CI checks that src-tauri/Cargo.toml agrees.

Built by

mert cobanov · 2026