Q: improving sound of old .mp3 files by filtering artifacts? Options

[Porcus]

Before shouting «no, you cannot know what has been deleted»: this is not about knowing, it is about guessing better than one could once upon a time in the previous millennium.


Back in the 90's, you could way more often than today single out an mp3 file by its artifacts (for given bitrate in suitable range [*]). Of course we could not know without the original whether a certain signal is an encoding artifact or is part of the music, but back in those days we could fairly often make an educated guess by listening -- doing «ABX except without the A and B».

So, the last 15 years have brought forth lossy encoders which are better at avoiding audible artifacts: Not only do we know better what to avoid, we also know better how to avoid it. And in principle, it might very well be possible to use this improved knowledge to (i) detect in an old mp3 file «what would have been avoided today», and (ii) filter it out. Not to achieve the impossible certainty that this part of the signal was not in the original source, but to achieve fewer audible nuissances more often than not.


Q: Has this been attempted, successfully or not?


(Any such tool could be easily testable if one has an old encoder and a lossless source and a few good ears.)




[*] I presume that http://listening-tests.hydrogenaudio.org/s...8-1/results.htm should keep TOS#8 moderation off my back.

[saratoga]

How often do you have three or copies of the same track each from a different encoding? Seems like even if this were possible you'd never actually have a use for it.

[Porcus]

No, I am talking of one. One old mp3 with artifacts we could by now know how to identify algorithimically and filter away in a way that more often than not results in an improvement.

[saratoga]

No, I am talking of one. One old mp3 with artifacts we could by now know how to identify algorithimically and filter away in a way that more often than not results in an improvement.

I assumed you mean A/Bing multiple samples to figure out what was an artifact. If not, how exactly do you propose to figure out what is an artifact? And what would you do about it?

[MichaelW]

I assumed you mean A/Bing multiple samples to figure out what was an artifact. If not, how exactly do you propose to figure out what is an artifact? And what would you do about it?

I take it that what Porcus meant was that some kinds of sounds are characteristically artefacts of bad MP3 coders, so that if he hears, say, ringing in an old MP3 file, he doesn't for sure know it's an artefact, but he'll take the chance and remove it, if that is possible. It might have been deliberately introduced--you could only tell for sure by comparing--but he'd still like it gone. Would that be possible?

[saratoga]

It might have been deliberately introduced--you could only tell for sure by comparing--but he'd still like it gone. Would that be possible?

How do you "remove" something without knowing what it sounds like? Aside from muting the audio that is

Its rather hard to "remove" something that occurs due to missing information or excessive approximation.

[nu774]

Processing filters to remove encoding artifacts such as blocking, banding, ringing are rather common in video.
Of course lost information cannot be recovered. They are applied just to get better perceived quality...though they usually have unwanted side-effect, too (removing noise usually removes not only noises but also details).

OP might be thinking of something like that... just a guess.

[Porcus]

How do you "remove" something without knowing what it sounds like? Aside from muting the audio that is

We know what it «sounds» like. I tried to explain that in fair detail in the original post, but again: even if we do not know for certain for every given sample that this particular property is an artifact or not, we know enough to guess better than 50/50 on real-world music (this for the relevant bitrange).

We also have knowledge about how to avoid it, given that we know the original. Again, not perfect knowledge, but fairly good knowledge -- this is the reason why you can encode a 1000 kbps FLAC file from CD into a 100 kbps mp3 file and still sounds ... if not transparent, so arguably superior to a random 90% removal of information.

And this knowledge has improved over the last 15 years. (Of course, Moore's law makes it harder to distinguish between improved knowledge and improved ability to utilize it.)

So then comes the question: has the science of psychoacoustics gotten us far enough to filter even without the exact original? We do not need to be able to filter everything -- it suffices that for a suitable percentage of low-bitrate files, it improves, and the rest it can leave untouched.

Processing filters to remove encoding artifacts such as blocking, banding, ringing are rather common in video
[...]
to get better perceived quality

Good point. Another example is the crackle/pop removal algorithms for digitalization of vinyl. (And the good'ole subsonic filter on phono pre-amps.)

[2Bdecided]

It's just an audio restoration challenge that (AFAIK) know one has really tackled yet. (IIRC creative claimed some of their sound cards did this, but I don't know if they really did).

When there's enough content that only exists as low-ish quality lossy audio that people want to re-issue in <tomorrow's new wonder audio format>, someone will come along and invent a process for making it sound better.


If it was me, I'd go after the splashy metallic artefacts that ruin the high frequencies first. Little idea how to though - most attempts to re-synthesize high frequencies sound pretty metallic themselves.

Cheers,
David.

[slks]

I think one very crude way to do this could be to just use a lowpass filter. Indeed, some of the most annoying MP3 artifacts come from poorly-encoded high frequencies. You could add a lowpass filter at, say, 7 kHz and remove all the high-frequency content, and your audio might sound less artifacted. However, this assumes there aren't also artifacts in the lower frequencies (which is often the case with poor encodes). And you'd also be able to tell very apparently that the audio had been lowpassed, it'd sound like something recorded in the 1950s. Whether or not that sounds better than leaving the artifacted high freqs intact is entirely subjective.

I don't know much about the theoretical side of audio processing, FFTs and such, but it seems like removing MP3 artifacts would be very difficult, if not impossible, and the results would not be that great. Doesn't sound like something that would be worth someone's time to develop.

[Gumboot]

For badly encoded voice you might try recompressing it with a speech codec to see if that is forced to throw away some artefacts it considers unencodable.