OpenUnmix

OpenUnmix is an open-source deep learning system for music source separation. It provides a reference implementation and pre-trained models that can separate a stereo audio mix into its component stems, typically vocals, drums, bass, and other.

Developed by researchers at Deezer Research and released as an open project, OpenUnmix aims to support reproducible research and practical separation workflows for music production and analysis.

Technical approach: OpenUnmix uses a neural network trained to estimate the magnitude spectrograms of each stem from the mixture’s time-frequency representation. The architecture resembles a U-Net encoder–decoder operating on magnitude spectra, reconstructing each stem by combining the estimated magnitudes with the mixture phase.

Implementation and usage: The model is implemented in PyTorch and distributed as an open-source package with a Python API and a command-line interface. It ships with four-stem pre-trained models and can be fine-tuned on datasets such as MUSDB18 or MUSDB18-HQ. It is designed to be accessible to researchers and developers and integrates with common audio processing libraries.

Impact: OpenUnmix has served as a widely used baseline in music information retrieval and as a teaching and experimentation platform, influencing subsequent open-source separation tools and tutorials.