Actually you have not. In order to use Ensemble Mode one needs to use GPU Conversion, if this is not possible (maybe because the lack of an NVIDIA GPU) then it is not possible AT ALL to use Ensemble mode. All of you have failed to make this clear to those who are new to this.

 

Last edited: Jul 8, 2023

And to add to that, modern Macs don't have NVIDIA graphics cards because Apple no longer supports them. So as long as no AMD graphics cards are supported, Mac users are screwed unless they use external solutions.

 

Which model is MDX23 v2/2.1 from this pull-down menu in UVR5 settings?



Or does this need to be loaded separately? Thanks!

 

Tried to install this using Linux via wine but I guess it needs CUDA or something. Nevertheless, all it did was crash.

However, The website service Fadr seems to work pretty well. Can rip a few tracks a day with a free account.

Does anybody do this stuff on Linux with a non-Nvidia video card? Looking for tips.

 

ohhh my god! this Colab is amazing!! https://mvsep.com/quality_checker/leaderboard2.php?id=4231 thanks for shared this!!

 

This is available only on Colab.
https://colab.research.google.com/g...23-Colab_v2/blob/main/MVSep-MDX23-Colab.ipynb

Original version by ZFTurbo can run locally, but it's not updated with last released models: https://github.com/ZFTurbo/MVSEP-MDX23-music-separation-model

 

You could have asked how to do when you have no GPU, instead of that rant, and I would have explained how to do... But no.

 

Last edited by a moderator: Jul 9, 2023

@jarredou Hey, I've been trying to use your amazing product. I want to say thanks for sharing this instrument, I feel that it's priceless.

Can you please share with me a quick way of downloading processed audio? I've never used collab; as I understand, it's best to zip all tracks inside collab, to somehow speed up, since Collab doesen't give too much kbps... What code should I use to quickly zip my output folder and then download it?

 

Best way is to save the processed audio directly on your mounted Google Drive, and then to download the files from Google Drive, it will be way faster than downloading them from Colab (and if you download the folder containing all the processed file, it will be automatically zipped before the download start).

 

need help ! ;) what are the best settings to separate a vocal from an intro interlude ? thanks in advance

 

I think Kim Vocal 2 is the best currently available model for vocals

 

Its actually the ensemble one with a few mixed models that has the best score on the board as we speak.

https://mvsep.com/quality_checker/leaderboard.php?sort=vocals

 

where can i get it? didnt see it in uvr5

 

@jarredou , I see that there's a new sheriff in town.

Can we use those in UVR5 ? MDX23C, Grow 128, Hop 1024

If not, where can we use those ? Are they available yet ?

Thanks for your time, always appreciated.

 

Last edited: Jul 16, 2023

You can use MDX-Net: 292, 496, 406, 427, Kim Vocal 1, Kim Inst + Demucs ft


In UVR5 it's under: Ensemble - Vocal / Instrumental
MDX-Net: Kim Inst
MDX-Net: Kim Vocal 1
MDX-Net: UVR-MDX-NET Inst Main
MDX-Net: UVR-MDX-NET Inst_full_292
MDX-Net: UVR-MDX-NET_Main_406
MDX-Net: UVR-MDX-NET_Main_427
Demucs: v4 | htdemucs_ft



MDX-Net info:
Batch Mode

Demucs info:
CLI version
Model: hdemucs_ft
Default options

Time Elapsed: N/A

Algorithm: Average / Average

 

Last edited: Jul 16, 2023

@jarredou I got one more question. Do overlapping and memory chunk size affect quality of the demixed audio? As if it were able to create artifacts and/or bleeding

 

These new models are only available on mvsep1.ru (who trained them) at the moment. The new ensemble is premium only.
UVR will publish new model with the upcoming update (currently in beta).

The mutlisong leaderboard is more complete and is based on real songs (the other is synthetic, mixing unrelated stems together to create a fake eval dataset) : https://mvsep.com/quality_checker/leaderboard2.php?sort=vocals

They do, chunk_size should be set to the maximum you can with the available memory ressources (song length can also influence memory usage), it's quite already at the mamimum in my MDX23-Colab default settings (for Colab free accounts).
Overlap is also improving quality, but has less influence than chunk_size. It aslo help reducing some artifacts happening at audio chunks boundaries.
Overlap=0.8 is good value for balance between quality and bearable processing time.

 

Last edited by a moderator: Jul 17, 2023

I have 588 credit still on mvsep1.ru.

I'll give it a shot.

If I understand correctly, the second Ensemble called ''MDX23C'' Ensemble is the one that is currently giving the best result for vocals. Do I understand correctly?

Thanks, I appreciate your help.

 

That's it !

 

Is there any model which can remove thousands of screaming fans from a live recording?