AI-based spectral estimation for pitch shifting?

[paulm12]

I'm curious how people deal with the fact that if you have a sample, and you pitch it down (say you sample a synth on C3 and then play A2 or something lower), linear processing will leave a gap in the frequency spectrum in the upper range. From what I understand, a lot of times people use phase vocoders or sinusoidal spectral modeling for this, however it can be difficult to have a general purpose model to fill in part of the missing spectrum. One way that I handle it is layering white noise or adding distortion (which generally adds 2nd order harmonics), however it seems like an AI-based method would be better able to analyze the current spectrum of the signal and generate the relevant harmonics based on the source material. Additionally, it would be cool if the algorithm could also find and possibly "unwrap" any aliasing that occurred when the sample was recorded at the given sample rate and is pitched down.
Are there any algorithms like this out there? Or how are you getting around this problem of missing spectra when multisampling/sampling?

[BertKoor]

You don't need AI.

Use 4 samples per octave, then repitching is one semitone at most.
Or sample with higher sampling rate.
Or only pitch up, never down.
It was a sample of a synth? Don't sample, use the original!

You don't need AI.