open-unmix-pytorch VS umx.cpp

In my first post, quite a lot of alternatives were discussed: https://news.ycombinator.com/item?id=36707877

The model I'm using is called Open-Unmix (https://github.com/sigsep/open-unmix-pytorch). In 2021, there was an update to Open-Unmix to include new weights, UMX-L, which made it perform better than it used to on the older weights (UMXHQ).

In the grand landscape of music demixing, I don't think UMX-L is near the top anymore.

_However_, the demixing performance of freemusicdemixer.com is very close to the full PyTorch performance of Open-Unmix UMX-L, despite the tricks I needed to get it working in the browser, such as splitting up the inference to operate on segments of the song, or making the LSTM operate on streaming segments rather than holding the entire track in the LSTM memory.

In my first release, I loaded and did inference on the entire track at once (like the PyTorch model), which frequently crashed or exceeded the 4GB WASM memory for medium or large-size tracks.

* Post-processing step (bigger impact)

I tried to tackle the post-processing step in my C++ code (which would win ~1 dB in quality across all targets) but it's too tricky for now [2]. Maybe some other day.

1: https://github.com/sevagh/free-music-demixer/blob/main/examp...

2: https://github.com/sigsep/open-unmix-pytorch/blob/master/ope...

Or open-unmix. I've found it performs better than spleeter in some of my tests.

There's maybe something useful for you here?

https://github.com/facebookresearch/demucs

https://github.com/sigsep/open-unmix-pytorch

https://github.com/bytedance/music_source_separation

https://github.com/deezer/spleeter

OpenUnmix ( is a neural network solution from Yuki Mitsufuji and Stefan Uhlich, music industry luminaries that work in Sony's core divisions.

Good question! So, I wasn't even thinking about WASM to begin with. When I saw llama.cpp and whisper.cpp on the front page of HN, I found the idea exciting - instead of neural networks being magic, I wanted to copy the ggml idea of parsing the PyTorch weights file myself and rewriting the inference code in a lower-level language than Python (or, it's even more accurate to say PyTorch, since there is so much matrix heavy lifting e.g. broadcasting or reshaping that is done for you automatically).

That's when I wrote umx.cpp [1] (which is what this site is based on).

On an unrelated project, a friend of mine mentioned WASM, and as I looked into it a bit more I thought trying to compile umx.cpp to WASM would be a great idea, since I only use Eigen (which is a header-only library that only depends on std).

1: https://github.com/sevagh/umx.cpp