Unweave
A desktop app that unweaves songs into their stems with local AI. Based on the excellent Sound Razor by pedromtcosta.
Unweave was called SoundRazer Desktop up to version 1.2.2. Installing Unweave replaces the old version and keeps your songs, settings and downloaded components.
About
Unweave packages Sound Razor into a native Windows application with an enhanced, modern UI. It uses AI models (Demucs, MDX, RoFormer) to separate mixed audio into individual instrument stems - vocals, drums, bass, guitar, piano, and more. You can also split guitars into lead and rhythm parts.
Features
- Native Desktop App - Runs in its own window using pywebview, no browser needed
- Enhanced UI - Beautiful, modern interface with improved UX
- 100% Local Processing - All AI models run on your computer, no cloud uploads
- Multiple Formats - Supports WAV, MP3, FLAC, M4A, OGG, AAC and more
- Guitar Splitting - Separate lead and rhythm guitar parts
- YouTube Import - Download and separate audio directly from YouTube
- Job Queue - Queue separations, exports and imports while another one runs
- Real-time Mixer - Control volume, mute, and solo each stem independently
- Playback Speed Control - Slow down tracks without pitch change for learning
- Project Management - Organize all your separated tracks in one place
Installation
For Users (Recommended)
Download Unweave-Setup-x64.exe (latest release) and run it. The installer is tiny and works offline:
- Installs to Program Files (default:
C:\Program Files\Unweave) — requires administrator rights, so Windows will prompt for them
- Creates a Start Menu entry and registers an uninstaller; you choose whether to also add a desktop icon
On first launch the app window opens immediately and shows a one-time setup screen. You explicitly consent to download and install the remaining pieces — FFmpeg and the PyTorch/audio-separation stack (~1.6 GB total) — before anything downloads. They go into %LOCALAPPDATA%\Unweave, which the uninstaller cleans up. AI models (~500 MB) are then downloaded on first separation.
For Developers
# Clone or copy this project
cd unweave
# Create virtual environment
uv venv --python 3.11 .venv
# Install dependencies
uv pip install --python .venv/Scripts/python.exe -e ".[full]"
# Run the app
.\.venv\Scripts\unweave.exe
Linux
Unweave also runs on Linux (x86-64). Install it with the bundled script:
git clone https://github.com/um1chc5/unweave.git
cd unweave
./installer/install-linux.sh
unweave
It needs Python 3.11+, uv (or venv), and GTK 3 + WebKit2 for the native window
(python3-gi, gir1.2-gtk-3.0, gir1.2-webkit2-4.1 on Debian/Ubuntu). The
installer writes a launcher and a menu entry, and the app downloads FFmpeg and the
separation engine on first launch, just like the Windows build. See
docs/linux-install.md for the full guide, uninstall steps
and troubleshooting.
System Requirements
- OS: Windows 10/11 (64-bit)
- RAM: 8GB minimum, 16GB recommended
- Storage: ~10GB free space (for dependencies and models)
- CPU: Modern 64-bit processor (4+ cores recommended)
- GPU: Optional (CPU works fine)
Technology Stack
- Backend: Python with Flask-style HTTP server (inherited from Sound Razor)
- Frontend: Vanilla JavaScript, HTML5, CSS3
- Desktop Wrapper: pywebview
- AI Models: Demucs, MDX-Net, RoFormer
- Audio Processing: audio-separator, PyTorch, ONNX Runtime
Credits
Unweave is based on Sound Razor created by pedromtcosta. All core audio separation functionality comes from the original Sound Razor project. Unweave adds a native window wrapper, improved UI design, simplified installation, a job queue and more.
It also stands on the work of these people and projects. The app downloads the models and the separation libraries during setup rather than shipping them; the Mel-Band RoFormer code is included in app/sound_razor/vendor/roformer with its license.
AI models
| Model | Used for | Author | License |
|---|
Demucs htdemucs_6s | 6-stem separation (default) | Meta AI · Alexandre Défossez et al. | MIT |
UVR MDX-Net Karaoke 2 UVR_MDXNET_KARA_2 | Vocals + instrumental | Ultimate Vocal Remover · Anjok07, aufr33 | MIT (UVR asks apps using its models to credit UVR and its developers) |
Lead/rhythm guitar mbr_lead_rhythm_guitar_listra92 | Guitar splitting | listra92 · hosted by noblebarkrr | Not stated by the hosting repository |
Code and tools
| Project | Used for | Author | License |
|---|
| Music-Source-Separation-Training | Mel-Band RoFormer architecture (vendored) | ZFTurbo (Roman Solovyev) | MIT |
| BS-RoFormer | Original RoFormer implementation the above builds on | lucidrains (Phil Wang) | MIT |
| python-audio-separator | Downloading and running the separation models | nomadkaraoke (Andrew Beveridge) | MIT |
| yt-dlp | YouTube import | yt-dlp contributors | Unlicense |
| FFmpeg | Audio conversion and export | FFmpeg developers | LGPL / GPL (the Windows build setup downloads is GPLv3) |
| pywebview | Desktop window | Roman Sirokov | BSD-3-Clause |
| PyTorch, ONNX Runtime, librosa, NumPy, SoundFile | Running the models and reading audio | Their respective authors | BSD / MIT / ISC |
Song info and covers
Song details lookup uses MusicBrainz (core data CC0) and cover art from the Cover Art Archive, both run by the MetaBrainz Foundation. Cover images belong to their respective copyright holders.
License
MIT License - See LICENSE file for details.
Original Sound Razor is also MIT licensed by pedromtcosta.
