Removing clean solo from multitude of clean guitars

SpectraLayers
1

Recently I use SpectraLayers quite much to learn guitar solos. The problem is: most of these solos are played with a clean guitar sound with other clean guitars in the background playing. I noticed this challenges AI stem splitters. I found Lalal.ai is quite powerful for separation tasks, but it very often fails to separate such parts.

Here is an example from my recent project. Here I was lucky with Lalal.ai, it could extract the guitars quite well from a part, but as usual, it couldn’t separate the solo only, it took the other guitars too. Now, the problem was: even having the music slowed down, separated by Lalal.ai, there where quite many places in the solo that was not clearly audible. You could hear something is going on there, but had no clue what. Opening the result from Lalal.ai in Spectralayers was kind of annoying, because you couldn’t see the solo sticking out from the rest. Sometimes it is easy to identify stuff, that was not the case now.

So I used the only method I knew for this case, the hard way: taking the rectangular selection tool, selecting portions that I knew are in the level of the solo, listen, if I recognised it is the solo, I selected the harmonic selection tool, and by trial and error I selected a group of patches, and after a while, luckily I could reconstruct the solo. But only the main notes, as most of the time these are note pairs that were played, but if you know the melody note of the pair, you can figure the whole pair out if you grab the guitar and start playing.

My question is: can this be done somehow that is easier and more precise? I am still using Spectralayers 4, I know newer versions have new features like stem separation, however when I tried the demo, that feature did not seem to be smarter than Lalal.ai. So I saw no reason to switch to a newer version for my needs. But maybe I missed something.

Let me show the example, so this is the song, slowed down with 11%, because I noticed, at the original tempo, it is imposible to catch some parts in the solo, it is so fast. Then the vocal was removed by Lalal.ai

And this is my project in Spectralayers. The green is the output from Lalal.ai, this is what it presented after extracting the solo (+other guitars), the blue is my selection with the harmonic selection tool, the stuff is copied to its own layer, so you can hear the solo more emphasized when playing the two layers together. Could this be done any better way with a material like this?

2

In more recent versions of SpectraLayers - I’m in SLP12 - you can do an Unmix Mid/Side.

Often - but far from always - solos are sort of concentrated to the Mid part of the stereo field. That is, after a Mid/Side unmix it might be easier for the ear to discriminate what you consider is the solo. It’s not easy here since the solo guitar is intimately interwoven with the accompanying guitars, and to my ears there is not any regular single string playing but more of a varying of the voicing of the chords where (mostly) the top notes make melodic lines.

You could then try to Imprint/Cast the Mid layer with the Sides layer as source (set to 1 px both horizontally and vertically) to bring down the distracting elements somewhat. Or, you could do an Umix Levels with Peak Type; with the threshold ‘round -12 dB, horizontal peak @ >35 px and vertical peak @ >20 px, for the same reasons.

Really nice and grooving piece!

3

Unmix chorus is worth trying. Even if you only have a single voice, then unmix chorus can take out the odd bit of left over guitar.

4

Thanks for the ideas. Indeed, playing with the middle / sides is a good idea. I know already that the mixing of the CD where this track is from is suitable to do it. Let me check the options you mentioned and we’ll see what the results will be