Open-Unmix UMX-HQ (ONNX export by Rostrum)
Open-Unmix UMX-HQ, a music source separation model that splits a stereo mix into vocals, drums, bass and other. This file is Rostrum's ONNX export of the four UMX-HQ networks (opset 17, any length); the STFT and the multichannel Wiener filter around the networks run outside the file. Compared with Open-Unmix's own PyTorch umxhq() on 15 s of music, every stem agreed within 106–129 dB.
This repository is a pinned, public copy of the exact files the Rostrum apps download. Rostrum fetches each file at a fixed commit of this repository and checks its SHA-256 before use.
What Rostrum uses it for
Sound's Clean up: it splits a mix into stems and pulls a voice away from music. It only separates what is already in the recording.
Rostrum uses it for analysis only, never generation: it reads what is in a recording or a picture and makes no new sound, voice or image.
Files
| File | Bytes | SHA-256 |
|---|---|---|
umxhq.onnx |
142,496,426 | b26b5f5a4ff0e919825d28b355bac648908a271f00990bf8ca2f238c87863ec9 |
In Rostrum's models folder they go under umxhq/.
Source
| What | Where | Commit / version |
|---|---|---|
| Original code, sigsep / Inria | https://github.com/sigsep/open-unmix-pytorch | fb672c9584997c2b05e148eeaa65b4c23ed4693b (master when mirrored) |
| Original weights (UMX-HQ, MIT) | https://zenodo.org/records/3370489 | Zenodo record 3370489 |
Authors
Fabian-Robert Stöter, Stefan Uhlich, Antoine Liutkus, Yuki Mitsufuji (Open-Unmix, Inria / Sony); ONNX export by Rostrum.
Licence
The UMX-HQ weights are MIT (Zenodo record 3370489) and so is the Open-Unmix code (LICENSE); our export keeps the MIT licence.
Training data note: UMX-HQ was trained on MUSDB18-HQ, whose own terms are for academic use. The weights themselves are published under MIT by their authors.
Citation
The original authors ask that you cite:
@article{stoter19,
author={F.-R. St\"oter and S. Uhlich and A. Liutkus and Y. Mitsufuji},
title={Open-Unmix - A Reference Implementation for Music Source Separation},
journal={Journal of Open Source Software},
year=2019,
doi = {10.21105/joss.01667},
url = {https://doi.org/10.21105/joss.01667}
}