Skip to main content ↑ ↓

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: Are there AI tools to undo DR compression or to clear up low res audio?  (Read 3412 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Are there AI tools to undo DR compression or to clear up low res audio?

Hello!

Seeing AI tech used to some remarkable things like changing the vocalists in musical numbers, I wonder if there are any solutions to do other things. Like undoing all the loudness war dynamic range compression nonsense or to (credibly) restore the fidelity to low resolution audio, like some 4bit 22kHz LDPCM sountrack I have of an old game.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #1 –
Unlikely I guess. AI tech has been used for the remarkable things you listed, but it tends to produce low fidelity output, when it's OK at first glance, but not so much when carefully listening in a quiet environment. Where it would introduce defects that may be more noticeable or more annoying, than those from dynamic compression/clipping.
Especially not credibly. It's not going to be possible to verify credibility. A destroyed record leaves no information about the process used to destroy it, it's only possible to infer it in simple cases, for example when it's simply straight digital clipping (for which AI is not needed, there are combinations of "classic" algorithms that can reasonably try to improve from that). Which means, AI would have to guess that too. And no one would be able to tell if it guessed correctly. Unless someone can compare against the undamaged original, but then there's the obvious question, why tf would then not simply publish the clean original version and skip all this nonsense.
With low resolution audio, it's kind of similar, there are potentially infinite ways of downsampling (rounding, harmonic noise prevention, antialiasing filters...), but the space is probably smaller than with even more infinite possible ways of destroying a record with dr compression. Maybe if it's known for a given game, what downsampling routine was used for it, then an ML based algorithm could be trained on it (with other similar material where hi-fi source is available) and then one can hope for some reasonable output.
a fan of AutoEq + Meier Crossfeed

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #2 –
Quote from: magicgoose on
why tf would then not simply publish the clean original version and skip all this nonsense.

That's billion dollar question. What's a fact is that they virtually never release good masters anymore, for popular music. Only compressed gunk.

I figured if an AI system is specifically designed to recognize instruments and music. It could selectively change the volume of separate instruments (not necessarily just real ones) and notes based on a wast sample of properly mastered music. And not just using that, but cues from the music worked on itself, since the sound changes for a not played or sung on higher volumes. (Outside many synthetic stuff I guess)

I hypothesized something similar for low resolution audio. Where an AI could be trained to know how an instrument is supposed to sound like, then filter out the noise, or more like fill out the missing information.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #3 –
At least Paul McCartney lately anounced to release a song from John Lennon created from a demo tape thanks to AI. How i understand it they used the AI to seperate and restore the voice.
For general one knop huggle sound i bet MQ-AI is already in the works to start another ground breaking de-smearing licensed format.
Is troll-adiposity coming from feederism?
With 24bit music you can listen to silence much louder!

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #4 –
Quote from: Wombat on
At least Paul McCartney lately anounced to release a song from John Lennon created from a demo tape thanks to AI. How i understand it they used the AI to seperate and restore the voice.
Interesting.

Quote from: Wombat on
For general one knop huggle sound i bet MQ-AI is already in the works to start another ground breaking de-smearing licensed format.
I'm not sure I understand much of this sentence.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #5 –
This topic has been mentioned many times even within HA.
https://hydrogenaud.io/index.php/topic,112548.0.html
https://hydrogenaud.io/index.php/topic,122986.0.html
https://hydrogenaud.io/index.php/topic,93853.msg1007172.html#msg1007172
https://hydrogenaud.io/index.php/topic,108864.msg948350.html#msg948350

MQA "get caught" because the naive upsampling approach can be easily identified via very basic measurement methods. In foreseeable future there are always possibilities of high quality audio deepfakes, but then the already quite successful AI technologies at the current moment are already posing risks to many people. Think about the Hollywood strikes, as well as AI copycat of Anime/Manga artworks and so on, and other things like face/voice swap frauds, spambots, automated fake news, content farms and such.

It is quite natural that corporates or criminal organizations are more willing to invest in profitable things, and it seems that research about restoration of low quality ADPCM game music is not as profitable as the things mentioned in for example, the previous paragraph.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #6 –
Quote from: magicgoose on
then there's the obvious question, why tf would then not simply publish the clean original version and skip all this nonsense.

It doesn't necessarily exist anymore?
https://en.wikipedia.org/wiki/2008_Universal_Studios_fire

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #7 –
DR compression destroys source audio, so what you're asking for is impossible.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #8 –
Quote from: Porcus on
Quote from: magicgoose on
then there's the obvious question, why tf would then not simply publish the clean original version and skip all this nonsense.

It doesn't necessarily exist anymore?
https://en.wikipedia.org/wiki/2008_Universal_Studios_fire
This would have been much more disastrous if happened in 1988 or earlier, but ADC technologies in late 1990s are already capable of affordable 96/88.2k sampling with more than 17 bits of dynamic range with distortion much lower than whatever combinations of studio tape machines and tape formulas at 15/30 IPS.
https://web.archive.org/web/20030416004609/http://www.pcavtech.com/soundcards/compare/index.htm

Event Gina above the gold line for example was released in 1997.
https://www.soundonsound.com/reviews/event-electronics-gina
Stereophile measurements of the CardDeluxe which is also above the gold line.
https://www.stereophile.com/computeraudio/280/index.html

With advancements of restoration technologies there is even no need to rely on things like bias tone for restoration.
https://youtu.be/qqK6wgsh3QA
There are of course things like head azimuth alignment which should have been done before digitization, but can also be corrected digitally after digitization. Also common sense that VU meters are dangerously unreliable for digital metering but anyone can notice that the captured waveform is clipped or not by looking at the screen of a DAW.

So the unadulterated flat digital transfer itself can be considered adequate for re-issue of old analog masters in digital formats. All the loudness war and unnecessary EQ, plugin etc are just postprocessing and they are not embedded in the digitized audio data in the first place. Of course, to make more profit one can also release the restored digital master on vinyl or analog tape again like what Mo-Fi did.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #9 –
re. master availability, yes, for the vast majority of us, it makes no difference if the master is destroyed.
would be nice though, if you could hire professional ultra-trained underground musicians, who can take a request to "replay+rerecord X perfectly but get rid of all the flaws and optionally take Y, Z, ... into account" and deliver.
and maybe AI can help somewhat with the education and training effectiveness.
I've heard, even voice training already can allow a person to employ a surprisingly large range of voices, it sounds like magic and impossible, but apparently they really can do it.
a fan of AutoEq + Meier Crossfeed

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #10 –
Quote from: birdie on
DR compression destroys source audio, so what you're asking for is impossible.
Um. No. That's where the AI part comes in.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #11 –
Quote from: mzso on
Um. No. That's where the AI part comes in.

Citations (examples) needed.

Next you will say that AI can restore a bad 128kbps mp3 rip and turn it into a rich source original. Not going to happen unless you're talking about recreating audio, akin to image upscaling.

The problem with image upscaling is that it doesn't work. You cannot restore lost details even with AI. AI can fill in / invent lost details but the result will be quite different from the original.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #12 –
Something can be restored with deconvolution with specific scenario otherwise its black art.
Please remove my account from this forum.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #13 –
I'm optimistic that it WILL be possible...    Not to truly "restore" the audio because some information is permanently lost, but to "improve" the audio.

If we can listen to the music and analyze what "needs" to be done, the computer may be able to do it.   If the software can un-mix we could dynamically expand the individual instruments (and make other tweaks, etc.), and we should be able to "un-mix" any noise, and possibly any distortion, in the recording.

Quote
Next you will say that AI can restore a bad 128kbps mp3 rip and turn it into a rich source original. Not going to happen unless you're talking about recreating audio, akin to image upscaling.
If AI can create a song that sounds like Elvis, but that Elvis never recorded, it should be able to fix-up a bad copy or create an entirely new-better copy (or a new version) of an existing Elvis song.

Of course the copyright holders will try to stop it and the lawyers will love it!      So the question will be if we can get access to the AI tools & processing for personal use.


...But overall, I think the future will be creepy and annoying.    :D   And perhaps dangerous, depending on how much power/control we give to AI.



Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #14 –
Quote from: birdie on
Quote from: mzso on
Um. No. That's where the AI part comes in.

Citations (examples) needed.

Next you will say that AI can restore a bad 128kbps mp3 rip and turn it into a rich source original. Not going to happen unless you're talking about recreating audio, akin to image upscaling.

The problem with image upscaling is that it doesn't work. You cannot restore lost details even with AI. AI can fill in / invent lost details but the result will be quite different from the original.
If it can do something as complex as remove a persons voice from an audio stream and replace it with a synthesized voice of someone else sometimes (when it' done well) seamlessly, then the possibility of identifying and separately changing the volume of different instruments, with some declipping applied as needed seems positively simple.

Quote from: birdie on

Next you will say that AI can restore a bad 128kbps mp3 rip and turn it into a rich source original. Not going to happen unless you're talking about recreating audio, akin to image upscaling.


That's strawman throwing nonsense, and obviously impossible. However it doesn't need to recreate the original. Only identify the mangling of the mp3 encoder, and generate the missing info based on a large dataset the AI was trained on. The end result might be transparent at some point, as far as human hearing is concerned.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #15 –
Quote from: mzso on
If it can do something as complex as remove a persons voice from an audio stream and replace it with a synthesized voice of someone else sometimes (when it' done well) seamlessly, then the possibility of identifying and separately changing the volume of different instruments, with some declipping applied as needed seems positively simple.

That's strawman throwing nonsense, and obviously impossible. However it doesn't need to recreate the original. Only identify the mangling of the mp3 encoder, and generate the missing info based on a large dataset the AI was trained on. The end result might be transparent at some point, as far as human hearing is concerned.

Citations / working applications needed.

As for replacing voice, it's a much easier to solve problem.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #16 –
Quote from: birdie on
Quote from: mzso on
If it can do something as complex as remove a persons voice from an audio stream and replace it with a synthesized voice of someone else sometimes (when it' done well) seamlessly, then the possibility of identifying and separately changing the volume of different instruments, with some declipping applied as needed seems positively simple.

That's strawman throwing nonsense, and obviously impossible. However it doesn't need to recreate the original. Only identify the mangling of the mp3 encoder, and generate the missing info based on a large dataset the AI was trained on. The end result might be transparent at some point, as far as human hearing is concerned.

Citations / working applications needed.

As for replacing voice, it's a much easier to solve problem.
And where's your reference for this much easiness?
Working applications will, be available when someone makes one.

Re: Are there AI tools to undo DR compression or to clear up low res audio?

Reply #17 –
Voice extraction isn't an easy problem by any means, and the results speak (or sing) for themselves. Even with the state of the art tools (that must be demucs?) you can clearly hear defects on tracks which aren't near the bottom percentiles by "complexity" to separate.
But I've seen no good arguments for how voice extraction is harder than this thread subject, either.
a fan of AutoEq + Meier Crossfeed