Real mp3 upscaling?
May 20, 2015 at 4:23 AM Thread Starter Post #1 of 8

pogodrummer

100+ Head-Fier
Joined
Feb 27, 2011
Posts
117
Likes
22
Hello.
I know there's plenty of threads about this, but most of them are started by people who don't have a clue about how MP3 works and compresses things.

Now, i've been researching visual upscalers, such as the ones found in modern TVs, and also stuff like SuperResolution, which, by some algorithms, can guess intermediate pixels and make a picture much, much sharper.

Now i'm wondering, would this be technically feasible with MP3 files? Could the high frequencies above the hard lowpass be somehow guessed by an algorithm? 

And, given two separate rips of the same song, would it be possible to interpolate the two with a process similar to digital zoom in digital cameras, thus increasing the overall detail of the files?

I started wondering this while making edits for unreleased tracks featured on several Soundcloud mixes. Could it be possible to restore a resemblance of HF detail when the detail simply isn't there?

Thanks!
 
May 20, 2015 at 11:07 AM Post #2 of 8

RRod

Headphoneus Supremus
Joined
Aug 25, 2014
Posts
3,371
Likes
970
The thing is that hi-rate lossy codecs (e.g. 320mp3 or 256aac) are already at the point of being inaudibly different from their uncompressed versions for most actual musical material. It's a whole different ballgame than issues with low-resolution image sensors. In essence, with audio we already have the high resolution available, so we don't have to bother with techniques that "average" lower res versions.
 
May 20, 2015 at 11:22 AM Post #3 of 8

maverickronin

Headphoneus Supremus
Joined
Apr 5, 2010
Posts
7,390
Likes
420
I pretty much agree with RRod.
 
It should be possible to reconstruct some of the lost data, but modern lossy encoders are already so good that it offers very little benefit.
 
OTOH, there is probably a very niche market for that kind of thing.  I have all kinds of random video game remixes (some downloaded from $DEITY knows where, Napster? Kazaa?) in the early to mid 2000's that probably no longer even exist in any other format than a circa 2001 128kbps mp3.  It might help, for those.
 
There are products that actually claim to do this, but I've never really looked into them.
 
May 20, 2015 at 11:30 AM Post #4 of 8

RRod

Headphoneus Supremus
Joined
Aug 25, 2014
Posts
3,371
Likes
970
 
OTOH, there is probably a very niche market for that kind of thing.  I have all kinds of random video game remixes (some downloaded from $DEITY knows where, Napster? Kazaa?) in the early to mid 2000's that probably no longer even exist in any other format than a circa 2001 128kbps mp3.  It might help, for those.
 

 
But wouldn't you need multiple non-identical versions of those files to use these techniques?
 
May 20, 2015 at 11:48 AM Post #5 of 8

maverickronin

Headphoneus Supremus
Joined
Apr 5, 2010
Posts
7,390
Likes
420
  But wouldn't you need multiple non-identical versions of those files to use these techniques?

 
Theoretically you'd just need to know what kind of changes the encoder to the original content in order to reconstruct something that could have been the original.  The more details you had about the encoder, the more precise it would be.  That would require a lot of work with various versions of specific encoders and the more specific you got, the more cpu power it would take to do it, but it's at least possible.  If you had multiple copies encoded in different ways you would be able to compare them and reconstruct even more data.
 
Unfortunately, doing such a thing well would require a lot of work and probably not be worth the effort given it's limited application.
 
May 20, 2015 at 12:12 PM Post #6 of 8

RRod

Headphoneus Supremus
Joined
Aug 25, 2014
Posts
3,371
Likes
970
   
Theoretically you'd just need to know what kind of changes the encoder to the original content in order to reconstruct something that could have been the original.  The more details you had about the encoder, the more precise it would be.  That would require a lot of work with various versions of specific encoders and the more specific you got, the more cpu power it would take to do it, but it's at least possible.  If you had multiple copies encoded in different ways you would be able to compare them and reconstruct even more data.
 
Unfortunately, doing such a thing well would require a lot of work and probably not be worth the effort given it's limited application.

 
I meant more that you can't use these multiple-capture techniques in the OP if you don't have the multiple captures. Certainly some amount of reconstruction can be made from a single file, but like you said I'd expect the results to be kind of "meh", especially if we're talking about restoring content in the 18-20k range, which isn't terribly audible to most people anyway.
 
May 20, 2015 at 12:27 PM Post #7 of 8

ProtegeManiac

Headphoneus Supremus
Joined
Oct 29, 2009
Posts
16,319
Likes
3,080
Location
Manila
Now i'm wondering, would this be technically feasible with MP3 files? Could the high frequencies above the hard lowpass be somehow guessed by an algorithm? 

 
//////ALPINE receivers used to have the BBE MediaXpander feature, with settings from 0 to 3 on the CDA-9830 I had. The problem is that, at best, it will have to guess based on how severe the filtering is based on what it knows about MP3 compression, but it can't guess what the original file was actually like. None of the settings worked right on all files - I had 256kbps on all of them and MX2 would be just about right for metal but in some cases MX1 is enough to screw up a predominantly vocal recording (like a Best Audiophile Voices track getting some boomy twang out of some of the piano keys). It has no reference for how much is actually there on the original recordings but EQs the sound based on the MP3 filters,but if the track didn't have much that were shaved off in that range to begin with, they can be boosted beyond what it had.
 
From what little energy I put into video though, I think it potentially is a lot easier. It's not simply a matter of guessing or referencing the original, but more of how the video algorithm manages how the pixels display the image, basically assigning the additional pixels in a 720p panel to smaller areas in 480p media. From the ones I've tried out though its effectiveness isn't really going to make the images look identical (when you take screenshots and compare side by side), as in many cases it's even harder to see the difference when it's a fast movie (like an action sequence), but really more of "at least you won't get too frustrated using your DVDs on this 46in HDTV" instead of tossing out all the Criterion Collection DVDs an HT guy has. Not sure though if this has gotten better since then.
 
In any case, given the price of storage and data streaming, plus devices that can utilize lossless or at least 320kbps and many different kinds of storage media (like how cars can use 320kbps from thumb drives or even 2.5in portable HDDs), it's probably a whole lot easier to just use 320kbps or lossless. If anything, a streaming service for example can archive the custom reconstruction on each track they have since they have access to the lossless copy and can have the servers to store the profiles, but again from a business standpoint it will likely be easier if they just offered lossless and 320kbpsVBR files for streaming.
 
May 20, 2015 at 6:17 PM Post #8 of 8

pogodrummer

100+ Head-Fier
Joined
Feb 27, 2011
Posts
117
Likes
22
   
But wouldn't you need multiple non-identical versions of those files to use these techniques?


That was exactly my point. If all you could get were 128kbps rips of the same track, but the rips themselves weren't identical bit-for-bit, you could theoretically reconstruct something, at least that's what i meant.
 

Users who are viewing this thread

Top