Check out my website to order a free demo.: rockout:
https://AI-Lyd.no

Interesting! Can you add some highlight times I can skip to? Like head to head comparisons. And then when I have time perhaps what's the whole video.

Tl;DR; No, not all stem splitters are the same.

thanks for that.

I know having only used 3 or 4 of them that indeed they are not.

Yep, I tend to try a few different things, and use the best of each of them, depending on my tasks.

For me, Logic is *by far* the fastest and most convenient to use, so that gets used a lot for quick a dirty stuff. It's surprisingly good, and better than I expected it to be when it was pre-announced.

Other tools like UVR, RX, Spectralayers and Lalal.ai do a good job too, but not orders of magnitude better than anything else - you just typically get a different set of compromises/artifacts.

Hell, have you checked out Serato Sample yet? Its separation might not be the 'best' (RipX daw seems to work the best on material I have (admittedly briefly) I have checked it out on).... but just as a tool.... If I had had that 10-15-20 years ago etc....

No, I haven't tried that one...

It's more of an "instrument" obviously, as it is a sampler, but being able to pull the egg out of the cake on a loop, or full track within it, in your daw of choice etc... and edit points/reverse/tempo/pitch shift etc...

the ramifications in ease of use are quite impactful if you check out a review or two of it...

definitely worth knowing what it can do though in it's relative front facing simplicity, but point of production/work flow...

Yeah, the plugin is cool, I just meant I hadn't tried the stem splitter in it to compare it with anything else...

https://ai-lyd.no/demo/teddy-swims-t...r-a-i-demo-30/

SL11 comes out in 5 days. Should be competitive.

I think people tend to avoid RipX because it has a dumb DJ-like name, but it's funny that it's actually a bit smarter than SpectraLayers, which sounds much more sophisticated.

https://ai-lyd.no/demo/benson-boone-cry-a-i-demo-32/

You can compare SpectraLayers 11 vs RipX 7.10 and others here :
https://divideconcept.github.io/Unmix-Comparison/

Thanks. I think RipX is slightly better for my purpose, but YMMV. They are all still not imperfect anyway.

I can do better! :-)

Its reallly bad. Every STEM has bleed on it.

That's what you get for trying to demix Louie Louie by the Kingsmen

By the way, as you ARE ai-lyd, and what you do is manual and not software, I'll for sure be sending some work to decide for myself how well you do for some of my very difficult 60s mono demix snips.

You're the only entity I haven't yet tested....sl wins so far on all my work.

FWIW: I tried RipX with "Sympathy For The Devil" ... excellent results. Now that was of course for throwing an excellent remaster of the song to RipX. On the other hand some 45 year old cassette recordings from some rehearsals gave mixed results. But there were vocals, that could be got away with if you label them psychedelic and/or experimental. All in all I am extremely curious what will happen in the next 5 years ...

Hello. Just upload your track here, you're welcome

https://ai-lyd.no/en/order-form/
Robert
AI-Lyd.no

UVR is far and away the best, although documentation and workflow are lacking, so bit of a dark art. Information is out there about models and techniques, if you take the time to look. The secret seems to be multiple passes using a different, appropriate, model with each cut. Nothing else is even close when you get UVR dialed in.

Is there some online resource that at least gives some guidance in this "secret" process, for those of us who just want it to do as good a job as possible, without having to descend into a dark cave and sell our soul to black magic for some years before we can do it..?

That makes no sense. What source combination? Trombone+trumpet? Kazoo+MarshallPowerChords?

Which combinations of which modules for what source combo?

Even the github discussions veer to zillions of different uvr model combinations....with no particular combo getting a vote of best.

There is no best algo. Source combinations determine what one may try as the algo or combinations.....but there is absolutely no "best".

There is no one combination to solve all problems, and I didn't mean to insinuate there was. General knowledge can be acquired about the methods & models in UVR and then generally applied with some sense of what levers to pull. I'm not here to explain the basics of UVR between MDX, Demucs, or the different models trained within those. Of course you need to dial in the general application to specific material. How is that different from any other complex audio tool?

I have found the discussions on Github to be the most informative about UVR. This is a good one.
https://github.com/Anjok07/ultimatev...iscussions/689

It requires some trial and error, but having trialed and errored with UVR, and many alternatives, the bar for success in UVR is higher.

The Github discussions are the best online resource, e.g. see link in my last post.

And reddit is a surprisingly good source of information.
https://www.reddit.com/r/IsolatedTra..._vocal_remover
https://www.reddit.com/r/IsolatedTra...ings_questions

Trial and error was where I ended up once I had an understanding of which models to try and general workflows that come recommended. Most of the choices you are confronted with come down to the models to use. Still trying to make sense of the information here, and what SDR really means.
https://mvsep.com/quality_checker/synth_leaderboard

https://ai-lyd.no/benchmarking/spect...yd-comparison/

https://ai-lyd.no/demo/toto-africa-ai-demo-33/

"Africa" is not a difficult track to demix.

Do this Mike Brown track instead, including oboe/brass separation, the two string sections, find the piano& extract, clean bass extracted/isolated, both guitars demixed from each other, each drum cleanly on separate mono tracks ie...everything...etc.....

Haha.. :-)
I can do some of stem you mention. Is this a challenge?, Will you post your results?

This is on another level https://www.audioshake.ai/
As is the cost