Hi everyone,

I honestly never thought I’d see the day, but here we are. I’ve been tasked with removing vocals from an AI-generated track for a karaoke project. Apparently, the client liked the AI version better than the original. The only source available is a 128kbps MP3 ripped from YouTube.

I’ve spent the last two hours wrestling with RX12’s newly improved Music Rebalance and SpectraLayers 12’s Unmix Song module. Neither is cutting it, though to be fair, there is only so much juice you can squeeze from a low bitrate stone. As you can imagine, the results have been… well, horrendous is putting it mildly.

At this point, I’m ready to stop trying and just let AI fight its own battles.

For those of you with a paid LALAL.AI account: in your experience, does it handle vocal removal any better than the industry standards I’ve already tried?

How would you approach this if you were in my shoes with everything we have today, bar AI ? Is there a trick with the RX12 or SL12 ?

Let me know.

Here's the yt link :

For context, I hate SUNO or anything that created the track above. Just listen to the vocal air with the cymbals, that's where every stem separation algo failed. I understand that the most logical way to deal with this particular project is to say no, but not yet for me.

 

I fully admit not being a specialist on these things, but for the small vocal/instrumental separation duties I've found UVR aka Ultimate Vocal Remover with the MDX-Net "Kim Vocal 1" model to work wonders. It's free and runs locally so what can you lose? GPU processing is quick too.

I can't provide a comparison with RX12 but it's way better than 11 at least. There's a thread in here about UVR where there are a lot of information from people who know their stuff.

 

Last edited: May 2, 2026

I think I have read it here before, but dunno its name. Thanks, will look it up

 

Hey @Stevie Dude, thanks for confronting us all with a new AI problem and bringing us closer to the madness and the "new" issues. How about a simple "Sorry, that's not possible." You're not an alchemist or a magician. You can't turn copper into gold. I would clearly state that it doesn't work and decline the job.

 

MVSep does great job. And it´s free, though Premium account has some more functions, but the free reg is enough for me

 

UVR5 with additional models is good, but for some of them you have to find the beta version of UVR5 before you can download certain newest models. The Kim Vocal 1 or 2 are good, but that is really about getting a clean vocal. You might want to use BS Roformer to get rid of the vocal, or even just htdemucs_ft.

I think with a 128kbs mp3 you are going to end up with so much swirl, it might not even matter which model you pick. Any time you might spend trying to track down a better copy of the audio file might pay off better than the differences of individual separation models.

 

logic pro steam separations

here a download for the vocal
DOWNLOAD

 

Last edited: May 2, 2026

You'd figure at $1400 legit-retail (funkin' ridiculous) that RX12 Advanced should be able to do something better than freeware "AI" junk.
But Noooooooooo....



Or maybe as was mentioned already, a 128k mp3 is really not a great starting point for such an endeavor...

 

Last edited: May 2, 2026

Splitting AI generated music is harder than real music ironically, as the elements sometimes aren't seen as what they are - if that makes sense (sorta). There's in depth explanations on YT somewhere.

 

Give uvronline.app a chance. The free trial gives you access to decent models for voice and instrumental. Otherwise go with MVSep or UVR5 to do it offline. And may have nothing to do with it, but have you asked the owner of the channel if they could provide you the stems or higher quality files? Dude has like 5 different disclaimers to avoid copyright, so i don't think they would be against sharing them with you.

 

But declining jobs won't make you any money. IMHO it's a valid request from a customer. In the end the customer will get the version that is as good as it gets.

 

Haven't tested the RX12 separation yet, but my go-to would be UVR5. Under options, you can get it to auto-download extra models you pick.

For the task of getting just the instrumental (kill vocals) try this setup:

use Ensemble (Ensemble is a mode in UVR, it runs both models, and then combines the "best parts" they were able to extract, so longer processing but better quality in the end)
Select these 2 models:
MDX32C
UVR-MDX-NET-Inst_HQ_4

On my GPU the original UVR5 (stable) has stopped working, it hangs up if you enable GPU mode. UVR5 beta works though, and is able to download extra new models for you to play with. Can recommend going for that one directly to skip headaches.

UVR5 Beta (Win64) - https://github.com/Anjok07/ultimate...download/v5.6/UVR_1_15_25_22_30_BETA_full.exe

UVR extraction does work decently on 128kbps Mp3's off youtube.

 

Last edited: May 3, 2026

got this from mvsep: https://wetransfer.com/downloads/b1...ecef45cebbfa12d5ad7f6f3b20260503040918/2650f3

 

I think you just threw the baby out with the bathwater.

 

Right now, MVSEP blows the competition out of the water IMO.

I've been trying different services intermittently for the past few years. UVR was really great at first, but nowadays, while still better than some other options, doesn't sound so much superior anymore (btw Spectralayers is using UVR algorithms, if I'm not mistaken).

There's a song I use for stem separation, and so far I was frustrated cause every result has similar artifacts, incoherent instrument groupings (the guitar is sometimes on the same stem as the sax, the piano, or "other", totally arbitrarily).

I recently tried MVSEP. The free option sounded more or less like what I had done so far, but much much better.

However, when I bought a paid plan (prices are very reasonable), the results blew me away. I mean this thing separated each individual drum (from the master track) and it sounded somehow cleaner than UVR when I used the best drum kit extraction algorithms (separating the entire drum kit, not individual drums).

More than that, it not only separated the guitars and saxophone with absolute precision, it also created different tracks for when the guitar was drenched in FX (as opposed to, you know, the cleaner parts).

And here's the real kicker : it got better results on a 128 kbps mp3 of the song, whereas I fed a wav file to the other ones.

My advice is to give the paid option a try. It really doesn't cost much, and is totally worth every penny.

 

use mvsep and their latest BSRoformer algo.

 

mvsep hands down!

 

best poll ever

 

Ultimate Vocal Remover for me.
I'm sure there's better stuff, its just whats installed on my pc.