Hush

Cuts the pauses out of your video. A desktop window for auto-editor — no terminal, no Python, no internet.

Download · Русская версия

---

What it is

You record yourself talking. Half the footage is you breathing, thinking, and reaching for the mouse. Cutting that by hand takes hours.

auto-editor does it in seconds, but it lives in the terminal and has forty flags. Hush is a window around it. Drop your files in, pick how aggressive you want it, press one button.

The engine ships inside the app. Nothing is downloaded at runtime, nothing is installed, and it works with the network off.

Download

Grab the build for your machine from Releases:

File	For
Hush-macOS-arm64.zip	Mac with Apple Silicon (M1–M4)
Hush-macOS-x86_64.zip	Intel Mac
Hush-Windows-x64.zip	Windows 10/11, 64-bit
Hush-Linux-x86_64.zip	Linux, 64-bit

macOS will say the app is from an unidentified developer, because it is not notarised. Right-click Hush.app → Open → Open. Once.

What it does

• Drag and drop a queue of files, or add them with a button
• Three presets — Soft, Normal, Tight — for how hard to cut
• Padding around speech so words don't get clipped mid-consonant
• Silence threshold in dB, for noisy rooms
• Cut the silence, or speed it up ×2 / ×4 instead of removing it
• Speed up the talking itself, up to ×2, while the pauses go
• Cut by sound, by motion, or by both — motion detection keeps screen recordings where the picture moves but nobody talks
• Drop black frames — throws away fades and dead air
• Smoothing merges micro-cuts so the speech doesn't stutter
• Smooth transitions — a dissolve across every cut instead of a hard join
• Count without rendering — tells you exactly how much will go, in a second, before committing
• Normalize loudness on the way out
• Export to an editor instead of a file — DaVinci Resolve, Premiere Pro, Final Cut Pro, Shotcut, Kdenlive
• Russian and English, switched with one button, remembered between runs
• Update the engine from inside the app when you're online

Subtitles, offline

Hush also transcribes. Speech recognition runs on your machine — nothing is uploaded, and it works with the network off once the model is downloaded.

• SRT ready to upload to YouTube, or plain text
• One word per line — the bouncing-word captions Shorts are made of
• Translate into English while transcribing, so a Russian recording gives English subtitles in one pass
• Eight models to pick from, tiny (74 MB) up to large-v3 (2.9 GB). The quantised ones give you a big model at a fraction of the weight — turbo-q5 (547 MB) is the sweet spot: large-v3 quality for the size of small. For anything other than English, start there.

Downloaded once, the model is kept in your user folder and reused.

Finished video, or a timeline?

This is the choice that matters.

Finished video gives you an .mp4. Fast, done, no edits afterwards.

An editor gives you a project file — a timeline with the cuts already made. Nothing is re-encoded, so it's instant and lossless. You open it and fix the places the engine got wrong. For real work, take this one.

Target	You get
DaVinci Resolve	.fcpxml
Premiere Pro	.xml
Final Cut Pro	.fcpxml
Shotcut	.mlt
Kdenlive	.kdenlive

Your original is never touched

Results go into a Hush folder next to the source:

My videos/
  talk.mp4              ← untouched
  Hush/
    talk_no_pauses.mp4  ← result

The app refuses to write to the input path, and adds a number if a name would collide. You can point it somewhere else with Choose folder.

Honest limitations

The engine reads loudness, not speech. It has no idea what you're saying. That means:

• Background music defeats it. Music isn't silence, so nothing gets cut.
• Two people talking over each other comes out badly.
• Quiet consonants (s, f, sh) can get clipped — that's what the padding slider is for.

It shines on one person, one microphone, a quiet room. Talking-head videos, tutorials, podcasts, screencasts. That's the job it does well, and it does save hours there.

Build from source

git clone https://github.com/EmptyOverlord/hush
cd hush
pip install -r requirements.txt
python3 fetch_engines.py        # downloads the auto-editor binary
python3 hush.py                 # run it

To produce a standalone app:

python3 -m PyInstaller --noconfirm --clean hush.spec

The result lands in dist/. Cross-compiling doesn't work — build Windows binaries on Windows. CI does all four platforms on every tag; see .github/workflows/release.yml.

The icon is drawn by code, no image editor and no libraries — see make_icon.py.

Releasing

./release.sh

That's the whole thing. It bumps the patch version, syncs the version inside the app bundle, tags, pushes, and CI builds all four platforms and attaches them to the release.

Command	Result
./release.sh	1.0.0 → 1.0.1
./release.sh minor	1.0.1 → 1.1.0
./release.sh major	1.1.0 → 2.0.0
./release.sh 1.4.2	that exact version
./release.sh -n	dry run — prints what it would do, changes nothing

On a Mac you can also just double-click Выпустить релиз.command in Finder.

Built with

• auto-editor 31.5.0 by WyattBlue — the engine that does the actual work. Public domain (Unlicense).
• tkinterdnd2 — drag and drop.
• Tkinter for the interface. No heavy UI framework, which is why the download is mostly just the engine.

License

MIT. See LICENSE.