Stem separation apps and algorithms - quality?

There are a few stem separation apps out there. Some are using Deezer's free Spleeter tech under the hood, others not.

I suspect that different types of music perform better on certain apps.

I tried a few while doing an edit of a track the other day, one stood out in particular, for that track. I'll add that in here later, but what are your experiences?

T.

 

Googled a bit, this reddit thread confirms my views, lalal.ai sounds the best by far. But their pricing model is based on volume, 30 tracks for 20 bucks, so it might become quite pricey very quickly.
I bought Xtrax Stems 2 way back, it's the one I'm most used to.
I'm just indulging in playing with the stems format now, been meaning to for years.
I have yet to see if per-item based conversions like lalal.ai's are worth it - it depends on how much I'd use it. And maybe the Xtrax files are "good enough", or something else without per-item cost.
I do feel there is lots of room for improvement in this field, still.

 

I use “let’s unmix” iPad app and it works really well if you feed it with wav/aiff. Also isn’t too slow for my iPad mini 4 (without Neural AI but not sure if this app gets use of that technology as Algoriddim’s one).

If you are on mac (I will recommend you iPad pro or M1 chip family if you are into this) there’s also nuo-stems app which seems fair priced for the features it brings.

cheers.

 

Where are you NI??

 

I have tried Neural Mix Pro and iZotope RX8 Music Balance so far.

Neural Mix Pro sounds like crap. I am really wondering how someone can get exited on it and willing to spend 50 bugs.
On the other side Music Balance sounds much more professional but is still far away from perfection. Depending on the material you can get sometimes good results and sometimes not. I did not try anything other so I am wondering how other solution would compare to it.

Since NI is having a partnership with iZotope there could come up with something cool. A new stem creator already including the possibility to split existing audio files.

We are all waiting on this

 

I did not know about "Let’s unmix" so I took a quick look. I got the separation super fast (iPhone 12) but if I compare it to Neutral Mix Pro and Rx8 Music Balance it would be on place 3. Of course its really low priced but you get a lot of artefacts and in all sounds muddy. For me still Rx8 send by far better but for my opinion in some cases not good enough.

 

I guess, it all depends on source song. Its stems are either well separable, worse separable or almost unseparable.
Depends on FX used, mix.

One may hava building made of big blocks, one may have building made of bricks, one may have monolitics buiding made nonseparable to reusable units....

The same with stems in song. Items may be separable, items may be heavily intertwined, so that usable separation is hardly possible. Despite SW used.

 

While some tracks are certainly not separable at all, there's a vast difference between software when it comes to separation. lalal.ai is by far the best automatic split, although that one just separates acapella from instrumental, it does not do four channel.
It seems that there are 3-4 common algorithms out there, most are open, and then some companies (like lalal.ai) have some custom stuff.

I purchased Xtrax Stems a few years ago, back then it was the absolute best at stem splitting. Since then competitors have popped up, and Xtrax is demonstratively worse than those that came after, even the newest version.

Before that, I was using Accusonus Regroover, and Xtrax blew that one out of the water.

I just started a trial with RipX, that's a desktop app that is a bit like melodyne, but focused on stems. That one is very deep, with it you could do things manually that an algorithm can't quite figure out by itself. But since you're working with bits and pieces pulled appart into 8 channels by this algorithm, even that route is probably quite hit and miss.

But I would not assume, Kubrak, that if one algo can't split a track well, it can't be done. And the percentage of tracks that suck in one splitter and work in another, is higher than you'd think.