Quatuor

🇬🇧 English · 🇫🇷 Français

A native macOS app for audio stem separation — isolate vocals, drums, bass and other instruments from any song using state-of-the-art deep learning, running entirely on the Apple Silicon GPU. Four stems: a quartet.

Features

• Drag and drop any audio file (MP3, WAV, AIFF, M4A), or pick one with Browse….
• GPU-accelerated separation via MLX on Apple Silicon (Metal).
• htdemucs_ft — Meta's fine-tuned Hybrid Transformer Demucs.
• Synchronised multi-track playback — all stems in lock-step via AVAudioEngine, started on the same host time.
• Mute / solo / volume per stem — instant karaoke: mute the vocals and sing along.
• Waveform view with live progress and click-to-seek, MP3 export (single stem or batch, VBR V2 via libmp3lame).
• Adaptive memory: the MLX Metal cache is capped at 70 % of available system memory; model and buffers are released after separation.
• English and French, following the system language.

Ready to separate	Separating

Install

Download Quatuor-<version>.dmg from the latest release, open it and drag Quatuor onto Applications. The app is signed with a Developer ID and notarized by Apple: it opens with a double click. It needs a Mac with Apple Silicon and macOS 26.4 or later.

The first separation downloads the model, about 670 MB, from Hugging Face into the app's cache; every later one runs offline.

To build from source instead (the Metal toolchain is required: xcodebuild -downloadComponent MetalToolchain):

git clone https://github.com/gwenn-ha-dev/Quatuor.git
cd Quatuor
make package      # build/Quatuor.app

Usage

1. Drop a song on the window. Quatuor shows its waveform, duration and sample rate.
2. Click Separate stems. The 4:32 song of these captures took about a minute on an M1 Pro.
3. Play the result, mute or solo stems, then export one stem with MP3 or all four with Export all.

New goes back to the drop zone; the separated stems are temporary files, deleted when you leave the result or quit.

How it works

The song is decoded by AVFoundation and handed to demucs-mlx-swift, a port of Demucs to MLX, which runs htdemucs_ft — one fine-tuned model per stem — on the GPU. Each stem is written as a 16-bit WAV in the app's temporary directory; playback schedules the four files on four AVAudioPlayerNodes started on the same host time, so mute, solo and volume act on stems that never drift apart. MP3 export encodes those WAVs with LAME.

Stem, StemKind and StemPlayer keep their names throughout the code: stem is the audio term for a separated track, and it stays the right word whatever the app is called.

Build

Command	What it does
make build	Release build, warnings are errors
make test	Run the test suite
make run	Launch the app
make icon	Regenerate Resources/AppIcon.icns
make package	Archive the app into build/Quatuor.app
make sign	Sign it with a Developer ID, notarize and staple it
make dmg	Build the notarized build/Quatuor-<version>.dmg
make shots	Retake the README captures (outils/captures.sh)
make release-check	Say whether build/ is publishable
make lint	Check compliance with the project charter
make help	List every target

Dependencies

• demucs-mlx-swift, on top of MLX-Swift, for GPU inference. The htdemucs_ft weights are downloaded at first use from the Hugging Face repository iky1e/demucs-mlx.
• LAME 3.100 for MP3 encoding (LGPL), built from source by make deps and linked statically.

Everything else is Apple frameworks.

The README captures use "Happy Happy HardPopCoreTastic" by Mesostic, dedicated to the public domain (CC0, Wikimedia Commons).

License

MIT © 2026 gwenn-ha-dev — see LICENSE.