How do you bring back an audio file from compressiong for more quality. - How to make something less bad.

Ya don't. Read the initial plane ride scene in Zendegi. You might be able to do a bit of tuning to make it subjectively sound better, maybe do stem separation to take it apart and put it back together, but at some point you're just going to have to pay someone to make a cover.
 
  • Agree
  • Like
Reactions: Dante Alighieri, PointOfNoReturnPoint, ssaneenah and 4 others
Vecr said:
Ya don't.
Click to expand...
Yep, because you can't replace what was taken out.

Why-do-we-need-a-lossy-format-when-we-have-lossless_png.jpg
Source (A)

This visual image of lossless digital music (in this case a CD) vs the 320kbps MP3 file says it all. Note the thinning between 16-20kHz and the brick wall at 20kHz with the lossy file, whereas the highs from the original CD are all intact.

That being said, this particular CD looks like a victim of the Loudness Wars, as evidenced by the brick wall at 22kHz.

The lower the bitrate and/or sample rate, the worse it gets.
 
Last edited:
  • Like
  • Informative
Reactions: Lord Xenu and The Repeated Meme
Squishie PP said:
What was said above used to be true, but now there are neural networks that can attempt to reconstruct a higher-fidelity signal. Here is an example of one: https://kuleshov.github.io/audio-super-res/
Click to expand...
That's what I was talking about with the stem separation, you would stem separate, run something like that, then mix it back together. It might sound better, but the original information is just gone.
 
  • Like
Reactions: Pee Cola
Squishie PP said:
What was said above used to be true, but now there are neural networks that can attempt to reconstruct a higher-fidelity signal. Here is an example of one: https://kuleshov.github.io/audio-super-res/
Click to expand...
I saw that then got linked to github, but then thats when it fell apart idk what i am doing...Thanks for the input but idk how to use this. Most thing's i am finding are plug ins for DAW's or an API, so I am pretty much lost until a GUI comes around.
 
Vecr said:
That's what I was talking about with the stem separation, you would stem separate, run something like that, then mix it back together. It might sound better, but the original information is just gone.
Click to expand...
Better than nothing, I guess. There is also a model trained for stem separation here: https://github.com/deezer/spleeter
PointOfNoReturnPoint said:
I saw that then got linked to github, but then thats when it fell apart idk what i am doing...Thanks for the input but idk how to use this. Most thing's i am finding are plug ins for DAW's or an API, so I am pretty much lost until a GUI comes around.
Click to expand...
Theres a cloud-based utility here that will let you do something similar: https://replicate.com/nateraw/audio-super-resolution

But of course it costs money :c

This is sadly one of the crappy parts of the machine learning community's over-reliance on Python; it makes it far too difficult for end users to use most of the software created in this space.
 
  • Like
Reactions: Pee Cola and PointOfNoReturnPoint
Squishie PP said:
Better than nothing, I guess. There is also a model trained for stem separation here: https://github.com/deezer/spleeter

Theres a cloud-based utility here that will let you do something similar: https://replicate.com/nateraw/audio-super-resolution

But of course it costs money :c

This is sadly one of the crappy parts of the machine learning community's over-reliance on Python; it makes it far too difficult for end users to use most of the software created in this space.
Click to expand...
Sweet thanks! I really only need a Gui to help do this more than anything, I don't code anyway.
 
  • Feels
Reactions: Pee Cola
lossless codec actually isn't lossless overall since ADC's aren't perfect. 44kHz is just 2 samples at 20kHz frequency, not a lot at all.
Master audio gets taken in at 192kHz and 44 is just "Peasant grade" music
 
  • Optimistic
  • Thunk-Provoking
Reactions: Pee Cola and Vecr
Dawdler said:
lossless codec actually isn't lossless overall since ADC's aren't perfect. 44kHz is just 2 samples at 20kHz frequency, not a lot at all.
Master audio gets taken in at 192kHz and 44 is just "Peasant grade" music
Click to expand...
Unless you plan on slowing it down there's not much loss if any at all. It's not perfect, but it's not anywhere near the problem OP has. If it was just ADC issues, I wouldn't be able to hear anything wrong with his song on my (somewhat terrible) setup. I can hear the problem all right, so it's much worse than that.

Dawdler said:
It's actually pretty horrible as 2 samples are just enough for rough approximation and nothing more
Click to expand...
It really isn't. Maybe you have a super fancy DAC that really runs at that rate (and won't just smooth everything over with capacitors), but on both of my mid-end DACs I can't tell anything wrong with the quality on a well encoded song. This song, on the other hand, is terrible.
 
Last edited:
It's actually pretty horrible as 2 samples are just enough for rough approximation and nothing more
 
Dawdler said:
lossless codec actually isn't lossless overall since ADC's aren't perfect. 44kHz is just 2 samples at 20kHz frequency, not a lot at all.
Master audio gets taken in at 192kHz and 44 is just "Peasant grade" music
Click to expand...
Dawdler said:
It's actually pretty horrible as 2 samples are just enough for rough approximation and nothing more
Click to expand...
Two points uniquely define a sine wave, which is why 44khz was chosen to begin with. Unlike bit depth, where there's a whole range of quality and "good enough" seldom is, sample rate is much more deterministic.

 
  • Like
Reactions: Squishie PP and Vecr
JohnnyG said:
Two points uniquely define a sine wave, which is why 44khz was chosen to begin with. Unlike bit depth, where there's a whole range of quality and "good enough" seldom is, sample rate is much more deterministic.
Click to expand...
They don't define a sine wave but you are approximating it and also defining the input as a mixed sine wave.
For any measurements 10 points per Hz are barely enough, going to 100 for anything meaningful like harmonics hidden in the sine.
With 2 points the result is same if you are imputing a triangle or a sine and you are completely lost on the phase of the sound wave.
Luckily for the sampling, most natural sounds are some sort of sine wave with harmonics on top so a lot of deficiencies are hidden.
 
On second thought, the audio super-resolution model is probably not the best model for this application, since the problem in this case isn't low resolution. I found a paper with several model architectures that actually do solve the problem of reconstructing heavily compressed audio here: https://arxiv.org/pdf/2207.01667

Edit: Here is a page full of examples of what their model can do: https://sonycslparis.github.io/restoration_mdpi_suppl_mat/
It is honestly pretty impressive.
 
Last edited:
  • Thunk-Provoking
Reactions: Vecr
You must log in or register to reply here.
Back