Mp3 quality restoration?

[missing]

Not to say I believe what I'm saying, but just to enter the discussion, didn't the OP imply that he wanted to research whether it might be possible to write software that would learn artefacts with a view to removing them.

First, you have to understand the problem, and I don't think that this is a neural one at all. The reason that there are artefacts is because of what is removed - that's what causes the 'phasing/flanging' sound. So unless you can replace what's missing, you can't stop the artefacts. All you can do is, in all probability, to make them worse. This is because you'd be adding something essentially incorrect to something essentially missing - compounding a felony.

Basically, the nasty noises aren't added ones at all - they are a reaction between the louder sounding bands and the missing information in the masked ones. And this changes dynamically according to the content. For instance, on music where for periods some sounds don't change, a lot of data is slung out (a la DCT), and the content of many bands is seriously depleted. But if the music suddenly has a lot of changing parts to it, far less is lost, which is why some parts of music suffer far more (when it's encoded at a low bit rate) than others. The other ever-so-sight (!) problem with this is that not all encoders behave the same way. Think differences between LAME, Fraunhofer, etc. And bear in mind that the only thing that's standardised is the replay, and you begin to get the idea... those different encoders all have significantly different artefacts. So whatcha gonna do? The noise is, in principle, caused by the same thing that causes all that noise when you apply too much NR - which is of course why you can't improve the sound of an MP3 with that, either.

So basically, in an MP3 you are listening to a comb filter effect that moves in frequency, and which also causes variable amounts of content-related subtractive distortion, leaving no trace of what was removed. This is, of course, why they all sound bloody awful, especially at low bitrates.

Theoretically, every signal can be recovered with a given model; but the chance of it is very, very small -- thus not practical.

More to the point, it may be correct, by chance, but you can't know that it's correct.

Paul

[missing]

First, you have to understand the problem, and I don't think that this is a neural one at all. The reason that there are artefacts is because of what is removed - that's what causes the 'phasing/flanging' sound. So unless you can replace what's missing, you can't stop the artefacts. All you can do is, in all probability, to make them worse. This is because you'd be adding something essentially incorrect to something essentially missing - compounding a felony.

The neural net I was proposing is not for 'removing' anything -- in fact in would add (interpolate, but complex, as a neural network obviously) sound where it should to get rid of the flanging (usually between the 2 channels).

but of course it would be an experiment and I have no idea how well it would perform in practice.

The neural net I was proposing is not for 'removing' anything -- in fact in would add (interpolate, but complex, as a neural network obviously) sound where it should to get rid of the flanging (usually between the 2 channels).

I still don't think you've understood what is happening... Interpolation, in any way shape or form cannot possibly form any part of a solution to this. You can only interpolate between correct points in the first place, whatever algorithm you choose to use - and you don't have them.

Well there is audio between the points (unless we're at the beginning or the end of the file), the 'flanging' has to come from something, whether it's lack of information or not. If it's lacking, an extrapolation or interpolation could solve it -- but of course, solving the flanging does not necessarily mean the audio is restored. Frankly I'm not even interested in the original wave. If the resulting interpolated/extrapolated sound has no flanging, it is good enough, even if it does not represent the original sound. For example, so-called DSP plugins available usually recreate the lost frequencies from the known ones. That does not yield the original wave (in most cases, unless the original is exactly the predicted one), but it does sound better and get rid of 'muffling' or loss of bass. Of course the algorithm used in these programs is simple to facilitate real-time usage -- but a neural network would be much more complex, and obviously not real-time, but it doesn't matter for me.

That is, it could get rid of the flanging by interpolating/extrapolating information from the already-present one (which contains the flanging) -- it will most probably not represent the original, but it might get rid of the artifacts which is fine enough.