A macOS app that cuts one audio file into the separate sounds inside it, using the silences between them as the cut points. Drag a file onto the window and every non-silent stretch becomes its own clip: audition them, drop the ones you don't want, export the rest as WAV.
It was written for sample libraries — a single take holding twenty dice rolls, footsteps or door slams, that has to end up as twenty files.
There are no dependencies, and nothing to install alongside it: reading and writing audio is AVFoundation, the envelope analysis is Accelerate. No ffmpeg.
The file is read once into memory and reduced to an envelope — one RMS value and one peak value per 10 ms window, taken from whichever channel is loudest, so a sound that only appears on one side is never treated as silence.
A stretch is silence when its level stays under Threshold for at least Minimum silence. Everything between two silences is a clip. Each clip then grows by Margin on both sides so it doesn't start and end on an abrupt edge, and anything shorter than Minimum clip is dropped.
The margin never lets two clips overlap: it is capped at half the silence next to it, whatever value the slider is set to.
| Control | Default | What it does |
|---|---|---|
| Threshold | −40 dB | The level under which sound counts as silence. Raise it for quiet recordings, lower it for noisy ones. |
| Minimum silence | 0.50 s | How long a quiet stretch has to last before it splits the file. Shorter values cut more. |
| Margin | 0.10 s | Silence kept at both ends of every clip. |
| Minimum clip | 0.15 s | Clips shorter than this are thrown away — it removes clicks and breaths. |
All four retune the split live, without re-reading the file: the envelope is already in memory, so dragging a slider re-cuts the whole thing as you drag. Clips you had deselected stay deselected.
Each clip is a tile showing its waveform, its number and its length.
- Click the waveform to play the clip, click again to stop. The bars turn from light blue to deep blue as it plays, so the tile itself is the progress bar.
- Click the circle in the corner to leave a clip out. Deselected tiles fade and their waveform goes grey.
- All / None select and deselect every clip at once.
Bar heights are normalised against the loudest peak in the whole file, not per clip, so a quiet clip looks quiet next to a loud one instead of being stretched to fill its tile.
Export asks for a folder and writes only the selected clips, as <source name> <n>.wav —
footsteps 1.wav, footsteps 2.wav, and so on. Sample rate, channel count and bit depth are taken
from the source file, so a 16-bit 48 kHz mono recording comes back out as 16-bit 48 kHz mono. The
original file is never touched.
Anything AVFoundation can read can go in: WAV, AIFF, CAF, MP3, M4A, and more. Output is always WAV.
- macOS 14 or later
- Xcode 26 to build (
actoolcompiles the Liquid Glass icon)
./build.shProduces dist/AudioSplitter.app. While developing, ./scripts/build_dev_mac.sh quits the running
app, rebuilds it, copies it to ~/Downloads and launches it again.
The split logic has no dependency on AVFoundation, so it checks on its own:
swiftc -O -o /tmp/check AudioSplitter/SilenceDetector.swift check/main.swift && /tmp/checkIt covers the threshold, silences too short to cut, clips too short to keep, a margin large enough to
make clips collide, a file that is all silence and a file with none. The checks use precondition
rather than assert on purpose — assert is compiled out of optimised builds and would pass without
running.
AudioSplitter/
AudioSplitterApp.swift window, drop zone, tiles, waveform
SplitterModel.swift file reading, envelope, slicing, WAV export
SilenceDetector.swift where the cuts fall — pure logic, no AVFoundation
Player.swift playback of one clip from the in-memory buffer
AppIcon.icon icon (Icon Composer document)
check/ checks for the split logic
scripts/ development build
The whole file is held in memory as 32-bit float — roughly 23 MB per minute of 48 kHz stereo. That is what makes re-cutting instant while you drag a slider, and it is fine for the takes this app is for; hours-long files would want block-by-block reading instead.
MIT.