how do you use all those ? MDX-Net: UVR-MDX-NET Inst 3 + UVR-MDX-NET Inst Main + Kim_vocal_1 + Demucs: v4 | htdemucs_ft - Shifts: 2 - Ensemble Algorithm: Avg/Avg ? i thouht you can only use one?

 

I've explained this on the previous page.

 

I2U¥ Has done many tracks and is in the process of a 6000 Track rip project that you could Demucs and split for use as the vocals are already removed Latest Example rips below but I2U¥ has done many 100s before the project currently.

Everything on youtube is shared as mp3 freely in the I2U¥ Instrumentals channel on Telegram

Youtube channel
https://www.youtube.com/@i2u26/featured

Telgram channel
https://t.me/+4mR1owB87b9kMmVl

I2U¥ Telgram Chat Group
https://t.me/+9_gm0Pj8ias4ZWVl

Latest Example rips

 

Last edited: Mar 19, 2023

Thanks for the heads-up on this, how are the results?
I cannot get the google collab to work, I keep getting a fatal error:

fatal: destination path 'drumsep' already exists and is not an empty directory.​

Is there some way to download the actual Demucs model, and place it in the models repository so that it can be called upon in Python? as in,
demucs -n drumsep <File>

If you have the demucs model file a link would be appreciated greatly, thanks!

 

@Dyslexicon Check the Audio Separation discord channel, there´s a lots of stuff (also model sharing, though not sure for drumsep as I´ve just found the channel)
https://discord.gg/P7FhQFH

 

I'm sharing the DrumSep model here:

From what I saw and heard with few tests, sometimes, the kick is mistaken as toms, or only parts of kick, but it's really easy to fix this manually, overall quality is promising. The dev told me the dataset was 7h of drums made with addictive drums (mixture + 0% bleed separate parts)

 

Last edited: Mar 19, 2023

Beautiful, thank you!!

So how to actually utilize this model - can I call upon it in Python like a regular demucs model? (ie, demucs -n drumsep <file>)?
I cant find where Demucs models are stored - would you happen to know the directory path on Windows?
Any other dependencies or hacks I need to do, to add it to an already functioning Demucs 4.0 installation?

I found where UVR 5.0 stores Demucs models:
(users\Username\AppData\Local\Programs\Ultimate Vocal Remover\models\Demucs_Models)
The model files, ".th" extension, are accompanied by a complimentary ".yaml" file - not sure if this is a necessary component to have in order to get them to be visible in the program, or to run at all.

Cant wait to try this as soon as I know how to install it

 

I've only used it on Colab using the 2 bash scripts from the github (with few path editing before), and got it to work. On the colab, it seems to be the standard demucs command line, with this drumsep model given as model arg

 

How is the quality of the drum separation algorithm? I have a few stereo drum tracks I would like to rip the snare and kick from. Could somebody explain an easy way to use it? Any way to use it in Demucs or UVR?

 

The best result for drums would be to use HTDemucs_ft.

Demux4 HT (htdemucs_ft, shifts=10, overlap=0.95)

If anyone know how to use MDX23 (MVSep 2023.03.21) then it would also provide an incremental benefit. ( Here maybe ? : https://mvsep.com/)

As we speak, Multisong dataset | Original stems | bass/drums/other joined is the one leading the board but I think its private as I don't know where to get it.

Drums Leaderboard:
https://mvsep.com/quality_checker/leaderboard2.php?sort=drums

UVR is the safe path with htdemucs_ft as we speak.

As my personal experience, while it is providing a very good result in comparison to anything we ever used before including Izotope and RipX, you will still need to cherry pick the ''cleanest'' drum hits.

Hope that help.

 

Thanks for the info Martel! Very helpful.

 

Mvsep's new MDX23 algo is available on the beta site: https://mvsep1.ru . You can ask a free premium trial to ZFTurbo while site is in beta, in mvsep's chan in the "Audio Separation" discord posted above.

I've posted the model in the quoted message, so you can use like any other demucs model, see my comment there for the command line: https://github.com/facebookresearch/demucs/issues/422#issuecomment-1501932895

 

Last edited: May 5, 2023

Hopefully these new models may be incorporated/ported to UVR or some other app or architecture that can be processed on one's local machine rather than through upload-only sites. I really hope the best tech isnt going to be hoarded by web-only services or infuriatingly cumbersome architectures like google collab

I'd prefer that Demucs continue to reign supreme simply because it's free, currently the best (unless they drop the ball) and can be utilized locally rather than upload-only or pay-to-play extortion services.

 

Given the .27 score increase between htdemucs_ft and MVSEP, I wouldn't worry too much. Its not like its going to make a better job everytime neither as it is still extremely song dependant. It's just nice that they are still going in the right direction.

I'm sure all the other algorithms will keep on getting better and better anyways.

 

Seeing its scores and after listening, I think MVSEP new algo in an ensemble algo with htdemucs_ft inside and mixed or chained with other (new?) models.

 

SDX 2023 Challenge ends today!
If any of the new models can be run locally on your own machine rather than web-upload/process - please advise how!

 

htdemucs_ft still seems to be the only one available for download. There's still no MDX23 in UVR5. Who knows, it might change later on.

 

For fun and giggles, I just separated 5ZA - K!LL B!LL which is a billboard top 10 as we speak.

Here's the UVR seperation using ensemble (FLAC INPUT):
MDX-Net: UVR-MDX-NET Inst 3 + UVR-MDX-NET Inst Main + Kim_vocal_1 + Demucs: v4 | htdemucs_ft - Shifts: 2 - Ensemble Algorithm: Avg/Avg:
https://1drv.ms/u/s!AtfvmPHCKb6a7CnyFwA5jBzn5c1p?e=DmwkYE


Here's MVSEP MDX23 (FLAC INPUT):
https://1drv.ms/u/s!AtfvmPHCKb6a7Ch7_ZWX0z7xumri?e=zeKsHX

MVSEP is currently online only and it is a Paid premium subscription base.

20$ USD = 1200 minutes of conversion file. ( 1 song of 4 minute divided in 5 stems = 20 credits ) ( this specific one was 2:34 and cost 10 credits for 5 stems)

Thoughts?

 

This is the score of the original stems, not an algo. It's only there as a maximum SDR score reference.

About MDX23, the last score posted by ByteDance SAMI team is high. But I'm more curious about https://gladia-research-group.github.io/multi-source-diffusion-models/ that should (finaly) be released at end of month, with better models trained on extended data since the prepint was published.

 

Are they going public or will it stay private like MVSEP ?

How come they didn't take part in the competition ?

I guess their acronym is arxiv if I follow the research link on the page. There's a lot of leak in their input mixture 1 and 2 on their website but it does sound full and it seems to provide less phasy artifact then others. I dont know if its because of the cheesy composition they used or of its really better.
Where will they release their version?