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.
It's super exciting to see people using open-unmix like this! I worked with the creator of the model to try to do the same as a university project! Our solution was... not great, to say the least, but I'm happy someone else managed to do it!
That's awesome! I first got in contact with Fabien-Robert (https://github.com/faroit) during the Music Demixing Challenge 2021, where the UMX-L weights were first unveiled.
We have since discussed my projects a couple of times! I even got the idea for a streaming LSTM from him.
I think music demixing in general owes a lot of thanks to Open-Unmix and co (https://github.com/sigsep), who have relentlessly been publishing open-source models and related code (source separation metrics, dataset loaders, etc.) for years, and who blew the industry open with their MDX 21 [1] and SDX 23 [2] AI challenges.
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.