In the linked sample you can hear 2 guitars playing, one more in the left, the other more in the right channel. I would like to separate them to hear the left side one only, but I was not successful. Is there any method to do this in good quality? There is some AI that can extract the right one quite well, but still some portions of the left one is audible, and if you highlight the separation and delete from the original, some data from the left one will also be gone, and it sounds bad.
You have an unrealistic expectation. AI models don’t listen like humans. We can instantly detect the two guitars, but they sound very similar, occupy similar frequencies and are melodically intertwined in complex ways. The playing technique for each part is also very similar. There isnt a clearly discernable rythym part and lead part, so how do you expect an AI model to detect and isolate the two parts with the accuracy you demand?
Yes, not AI but you can perform an Unmix Mid/Side, and then duplicate the Side Layer, and then use the Channels Remixer on each of those duplicates; one for the left and one for the right channel.
I believe this will yield better results than using Unmix Multichannel Content.
An excellent question! And as a retired software professional and experimenter with SpectraLayers, I’ve pondered that very question. I can clearly hear the two parts, so their “ought” to be a way to describe the algorithm to separate them.
Not quite that simple though, huh?
Even doubling of dissimilar instruments is a challenge but attempting to determine how to separate two stringed instruments playing in unison is quite the puzzle. Two guitars, even with micro tuning differences, still has to account for harmonics, bends, etc.
I’ve gotten stuck on stem separations where I can clearly hear, as well as see the differences on the spectrogram, and back comes that “ought”!
Given that, and assuming you were asked to help a programmer determine how to identify the individual parts, where would you start?
I don’t know. But meanwhile I could separate them by AI. Interestingly the AI itself couldn’t solve the task, in spite of that it could separate the lead guitar mostly precisely. But was not smart enough to use is as a selection and remove it from the summarized guitar extraction it did (rhythm and lead). So I did it instead. This will be suitable now to study the guitar parts. I double checked it in Spectralayers, and the separation does seem to be correct. The appropriate notes appear in the appropriate file.
I think in this case the automation would be simple: listen to the right and left sides, and pick the notes that are above a certain loudness. One guitar is panned to the left, one to the right. A minimum amplitude should be defined, above which it extracts the notes from that side.
that’s a result, but certainly not ideal.
it’s just a function of the task difficulty versus the technology capability, which is limited by training.
I mainly need it for learning, but it would be useful if I could play it along with the song. For the latter the extraction (removal) the way as it is might not be perfect, but still might do. If the boom of the kickdrum was snapped out that would be a bigger issue in case I was going for the drums. By the way, I noticed that it is the beginning of the transient that still persist from the other guitar in the extractions. I could improve this a bit by shifting the selection a bit backward, to come a tiny bit earlier than its corresponding part in the mix. That way when you press delete, those transients are further diminished, not completely though, but the result is better.
But I am thinking about an improvement, I just don’t know if it is possible to do it in Spectralayers? When listening you can hear the two are panned to different sides. Couldn’t the considered amplitude level be defined for the selection? Or a diminishing of amplitude until the more silent guitar part goes missing on its side? That way only the louder dominant guitar part would remain on that side. And it might solve the problem or would offer another possibility for a better more precise selection.