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 at 10:16 PM

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 at 10:33 PM

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 at 10:47 PM

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.