What's new
By:
Filters
LiteRECORDS

Register a free account today to become a member! Once signed in, you'll be able to participate on this site by adding your own topics and posts, as well as connect with other members through your own private inbox!

  • Guest, before your account can be reviewed you must click the activation link sent to your email account. Please ensure you check your junk folders.
    If you do not see the link after 24 hours please open a support ticket.

Demucs - Stems Extraction Tool

NewAgeRipper said:
Yes I read that. I'm just trying to find out if the extras model is already in the colab method. :p
Click to expand...

Yes, that's why there is "demucs_extra" in the code :)
If you want, you can try the regular model, just remove "_extra" from the name, but I highly doubt it will be better. The extra model was trained with more songs.

Also, there is another model that sometimes provides better drums and bass (mostly on songs with a lot of acoustic drums and guitars), it's called "tasnet_extra". Just replace that in the code where you put your song name, though it takes more time to finish than demucs_extra.
 
I personally found that the vocals on this stems collab works better than the usual AI that was originally created. It doesn't make instrumentals as well. But does do a good job isolating otherwise. It's insane how clear a lot of this stuff is.

Thanks to all that are working on this. I really appreciate it.
 
Does anyone know if it's possible to "train" it ourselves using the stems and multitracks here on LR?
 
djtayz said:
Yes. you can. You can read about it here https://github.com/facebookresearch/demucs
You will just have to replace the MusDB dataset with your own pairs of stems.
Click to expand...

Interesting, will have to take a further look.

Just tried chucking Harvester of Sorrow off of Metallica's "And Justice for All" album through it to see if I could extract the bass from it...unsuccessfully. I knew it was probably too much to ask, but they really REALLY buried the bass on that album. You get little blips here and there of it coming through which is amazing in itself to be able to pick that out of obscurity, but nothing concrete enough to hear the full bassline sadly.
 
Further to my previous post though, the other stems off of the run through of Harvester actually sound pretty decent, some slight leakage between the tracks occasionally, but this is amazing being able to do this to fully mixed tracks with no previously available stems!
 
Pearce said:
Further to my previous post though, the other stems off of the run through of Harvester actually sound pretty decent, some slight leakage between the tracks occasionally, but this is amazing being able to do this to fully mixed tracks with no previously available stems!
Click to expand...

You can try replacing "demucs_extra" with "tasnet_extra" on the fifth step. That model should work slightly better for metal tracks.
 
Excuse my question if it sounds stupid but when i'm in Colab and executing the steps, there is a lot of stuff downloading.
Where are the files going ? On my computer ? On my google drive ?
 
Angelman60 said:
Excuse my question if it sounds stupid but when i'm in Colab and executing the steps, there is a lot of stuff downloading.
Where are the files going ? On my computer ? On my google drive ?
Click to expand...

It all happens in the cloud dude. There is nothing to worry about. If anything were downloading to your computer you would see a window on the screen of some sort indicating that.
 
Thanks guys, and it's true that step 2 is a bit long like 12 minutes here. Is there a way to process 2 or 3 tracks in a row without repeating the first steps ?
 
Angelman60 said:
Thanks guys, and it's true that step 2 is a bit long like 12 minutes here. Is there a way to process 2 or 3 tracks in a row without repeating the first steps ?
Click to expand...

As long as you are active when you go through it the first time for the day, you can do as many songs as you want. You just need to be doing something every few minutes so it doesn't close the connection.
 
Angelman60 said:
Thanks guys, and it's true that step 2 is a bit long like 12 minutes here. Is there a way to process 2 or 3 tracks in a row without repeating the first steps ?
Click to expand...

If you already processed one track and didn't close the tab, you can just change the name to the other track and don't have to repeat the steps.
 
Comparing demucs with Ultimate vocal remover...

UVR is way better, for making instrumentals

Demucs:
I noticed a lot of remaining vocals in the "drums" stem.
And it seems the background vocals are left in the synth stem.
 
joaoasq said:
Comparing demucs with Ultimate vocal remover...

UVR is way better, for making instrumentals

Demucs:
I noticed a lot of remaining vocals in the "drums" stem.
And it seems the background vocals are left in the synth stem.
Click to expand...

Not in all songs. Background vocals are a must for me and that's even hit and miss. Just have to keep trying my friend.
 
NewAgeRipper said:
[MENTION=26971]djtayz[/MENTION] will we see a stacked model update for this one?
Click to expand...

There already is something similar, called "shifts", but it's not really a trained model, it's just slight variations in quality here and there and takes ten times more time to complete.
 
djtayz said:
There already is something similar, called "shifts", but it's not really a trained model, it's just slight variations in quality here and there and takes ten times more time to complete.
Click to expand...

Is there a command to change that can be typed in to try it out?
 
You must log in or register to reply here.