Arxiv is an open access platform for academic papers.

The one from Gladia ResearchGroup will be publish on their github space, there's a link to it in their demo site. As it's a diffusion algo, it does not extract data from the audio mixture but it recreate each stem (with more or less accuracy and artefacts), like diffusion networks create images from noise (with more or less accuracy and artefacts !)

For the ByteDance MDX23 one, the winner is said to open source its algo, so it should be available (and their algo from 2021 competition was made public after the contest)

EDIT - the official rules:



So in worst case, we'll be able to use it with their model, but they may not publish training code, so impossible to have 'community models' like we have with MDX/Demucs/UVR

 

Last edited by a moderator: May 9, 2023

In any case, this is really exciting. I love every single bit of it. But we came a long way from that izotope thing.

I tried asking RipX when they would be updating their seperation model and if they were using Demucs 4 and they told me they could not answer that question. ( i guess for legal reason !? )

I was trying to follow the board setups in UVR5 for an ensemble and realized that a few from the chain was missing like MDX-net 496, Demucs mmi and some Kim Vocal model ( not the 1 ) I couldn't download or locate in UVR5.

 

In UVR "Demucs-mmi" is "hdemucs-mmi" (it's demucs3 retrained with the 800 additionnal songs used to train demucs4)

"MDX-net 496" is "UVR-MDX-NET Main" (I'm not 100% sure, but 99%)

And the "Kim Vocal 2" is not public yet as it was used for the MDX23 competition, Kimberley should release it soon now that the comp is over.

 

Last edited by a moderator: May 9, 2023

You are simply awesome !!!

Thank you so much for sharing all of those info.

 

It's only the MDX23 algo that is premium access (and maybe future ones), all others are free, so it's still usefull for people who can't use UVR locally (and it's easier to use than a Google colab).

X-minus is a good option too, between all others paid options, it's one of the contributing dev of UVR who manages the site and also train new models that are accessible for premium users (weeks before being included into UVR most of the time)

 

Last edited by a moderator: May 10, 2023

I highly suggest anyone to encourage him / them if possible. The difference is negligible with the UVR Ensemble and the processing time is obviously a lot longer but the result is pretty much the same to my ears. Maybe a tiny bit fuller for MVSEP but that's splitting hair.

 

Can someone give me the TLDR of this thread and suggest the best AI stem separation algo?

 

Here's is the board with scores for all stems (and the models/algos used), you can sort results by clicking on column titles: https://mvsep.com/quality_checker/leaderboard2.php?sort=bass

Note that things should evolve in the next days/weeks as there was a "demixing competition", that is over now. The winner and some other participants will release their algos & models. The winner should be ByteDance, and their scores are higher than any other algos/models currently available. And some other new source separation method with be also published at the end of month (based on diffusion networks).

 

MVSEP's new MDX23 algo is released: https://github.com/ZFTurbo/MVSEP-MDX23-music-separation-model

 

Anxiously waiting for the run.bat to run the Standalone GUI.

 

Care to share another invite ?

 

MDX23 fine-tuned vocal model (MDX) from Kimberley Jensen is released :



https://drive.google.com/drive/folders/1ikAutU_SuskkQtbwcGFeFWkW6ElrmLn9

 

Here it is : https://discord.com/invite/V88KegPb7D

 

Here it is : https://github.com/ZFTurbo/MVSEP-MDX23-music-separation-model/releases/tag/v1.0

 

When running run.bat I get:

C:\MVSep-MDX23.zip>"./Miniconda/python.exe" gui.py
GPU use: 0
Use device: cuda:0
Model path: C:\MVSep-MDX23.zip/models/Kim_Vocal_1.onnx
Device: cuda:0 Chunk size: 1000000
Model path: C:\MVSep-MDX23.zip/models/Kim_Inst.onnx
Device: cuda:0 Chunk size: 1000000
Go for: C:/AI/DEMO/Demo/Elton John/SOURCE - Goodbye Yellow Brick Road.mp3
LLVM ERROR: Symbol not found: __svml_cosf8_ha


Any thoughts??

 

Better to ask support on MVSep's github space or in MVSep discord channel (invite posted above)

 

It crash both on CPU and GPU here.

Reported as an issue on Github:

https://github.com/ZFTurbo/MVSEP-MDX23-music-separation-model/issues/1

 

The model requires 11 GB of video memory, otherwise will default to CPU. Edit the bat file and run it with

Code:

"./Miniconda/python.exe" gui.py --chunk_size 200000 --single_onnx

This is how I managed to make it work here

 

That didn't work for me sadly. It still crashed at 20%

 

Remove the chunk editing in the bat file (and keep the single_onnx) and try this. In the inference.py file, line 244, set a lower chunk size as shown below. Also feed it a 10-secods long audio file, just to make sure it works. Then for a full audio file it will take about 2 hours. Then maybe increase the chunk until your computer start crashing.

I'm afraid with this new wave of AI tools we need to upgrade to 16GB video cards. ffs

Code:

        if device == 'cpu':
            # chunk_size = 200000000
            chunk_size = 1000000

PS: this setting work in my case on a 32GB DDR system. It constantly takes 30 GB of memory. If you have less memory you'll need to reduce chunk size even more

 

Last edited: May 13, 2023