- Home
- Stem Separation Software
- Open-Unmix
Open-Unmix
What it is
Open-Unmix is a free, open-source PyTorch reference implementation for separating music sources. Its ready-to-use models divide pop recordings into vocals, drums, bass, and other instruments. The project also includes umxse, a speech-enhancement model trained on the 28-speaker Voicebank+DEMAND corpus. Researchers can reproduce results, develop architectures, and add their own data; training is documented separately from the package. Pre-trained choices include umxl, umxhq, umx, and umxse, along with spectrogram variants. Each source model uses a three-layer bidirectional LSTM, and the project states that this architecture cannot operate online or in real time. Installation is available through pip or Docker, and models can load from Python or torch.hub. A Google Colab notebook supports experimentation without local setup. Command-line input includes WAV, FLAC, and OGG, but not MP3; stempeg can broaden formats and export STEMS. The software is MIT-licensed, while umxl weights allow non-commercial use under CC BY-NC-SA 4.0.
For and against
For
- Separates vocals, drums, bass, and other instruments
- Includes speech-enhancement model
- Supports pip installation and Docker
- Offers Python, torch.hub, and Colab access
- Allows batch processing and WAV export
Against
- Cannot run online or in real time
- Command line does not support MP3
- umxl weights restrict commercial use
- No desktop app is listed
The verdict
Open-Unmix provides free source-separation models and research-oriented extensibility across Linux, macOS, Windows, and self-hosted setups. Check its format limits and non-commercial umxl weighting terms before adopting it.
Who it suits
It suits researchers, audio engineers, and artists working with source separation or speech enhancement. It also fits users comfortable with Python, Docker, command-line tools, or hosted notebooks.
Plans and prices
Open-Unmix
MIT-licensed software · umxl model weights are licensed for non-commercial use
Read 8 Oct 2026 · github.com
On the record
- Audio formats
- The command line supports WAV, FLAC, and OGG files but not MP3; stempeg can expand supported input and output formats, including STEMS export.github.com · read 8 Oct 2026
- Installation
- The Python package can be installed with pip, and the project also provides a Docker container.github.com · read 8 Oct 2026
- Integrations
- Pre-trained models can be loaded from Python or through torch.hub, and a Google Colab notebook offers online experimentation without local setup.github.com · read 8 Oct 2026
- Maker
- The SigSep GitHub organization describes itself as providing open resources for audio source separation.github.com · read 8 Oct 2026
- Model architecture
- Each source model uses a three-layer bidirectional LSTM, and the project states that this design cannot run online or in real time.github.com · read 8 Oct 2026
- Model options
- Pre-trained options include umxl, umxhq, umx, and umxse, with spectrogram variants also available.github.com · read 8 Oct 2026
- Model weights license
- The umxl weights are licensed for non-commercial use under CC BY-NC-SA 4.0.github.com · read 8 Oct 2026
- Music stems
- Its ready-to-use models separate pop music into vocals, drums, bass, and other instruments.github.com · read 8 Oct 2026
- Purpose
- Open-Unmix is a PyTorch reference implementation for music source separation aimed at researchers, audio engineers, and artists.github.com · read 8 Oct 2026
- Security and compliance
- The opened project pages state an MIT license but do not describe security controls or compliance certifications.github.com · read 8 Oct 2026
- Software license
- The repository lists the software license as MIT.github.com · read 8 Oct 2026
- Speech enhancement
- The project provides an umxse speech enhancement model trained on the 28-speaker Voicebank+DEMAND corpus.github.com · read 8 Oct 2026
- Support
- The project directs users to GitHub issues, Gitter chat, and Google Groups for technical help and support.sigsep.github.io · read 8 Oct 2026
- Training and extensibility
- The code is designed to let researchers reproduce results, develop architectures, and add their own data, while model training is documented separately from the package.github.com · read 8 Oct 2026
