[giantrobot]

Audio compression literally throws out data. It's audio the human ear isn't likely/able to hear but it's still gone. Different codecs and even different implementations of codecs throw out different audio data.

Any sort of upscaling is going to involve a lot of guesses about what was thrown out. For every codec and implementation there will be a lot of possibilities of what the original data might have looked like. At the end of the process you likely aren't going to get a result that necessarily sounds better than the low quality encoding. There will likely be a lot of artifacts introduced by the upscaling.

It's not impossible but something you'd need a specially trained AI to do I would think.

[lostlogin]

I work with a system which extrapolates data in the frequency domain to reduce noise in medical images and then doubles the images resolution in the x and y direction (so 4x the pixels). It’s absolutely amazing.

It’ll be interesting to see if audio ever gets a similar treatment and in what areas it can be applied.

[PaulDavisThe1st]

Noise levels in contemporary audio is vanishingly small, and has not really been the target of any audio technology for quite a long time.

Of course, restoration of older recordings is a separate issue, and that technology is 20-30 years old, at least (and very, very good).