As a computer tech and audio enthusiast, I have been interested in the so-called "Artificial Intelligences" (actually machine learning and transformers, nowdays) and I have collected some interesting free/open resources for the HyMPS project in the AUDIO section \ AI-based page.

Among these, in the Enhancers subsection, I had collected some self-called "restorers" that have - through statistical inference - the goal to regenerate close-to-lossless from compressed audio files (e.g. MP3).

For some time I was relatively uninterested in them (in the meantime I helped an Intel engineer choose which upscaler to bring into their OpenVINO™ AI Plugins for Audacity) but recently I discovered that Apollo models has obtained quite interesting qualitative results:



This model - named MP3 restore - has been trained to "delossify" < 128Kbps sources and can be tested here:

- https://huggingface.co/spaces/patriotyk/Apollo
- https://colab.research.google.com/g...lob/main/Apollo_Audio_Restoration_Colab.ipynb

...it would be very interesting to organize a so-called "blind test" (in addition to a null one of course) to fully understand its effectiveness in relation to various musical genres.

 

I always admire people explaining new things. Unfortunately I don't have a clue what this is about, but thank you for trying!

 

some of my favorite music was only ever released in 128 (or 24 even ) so this is interesting to me, thanks!

 

Will I live long enough to witness Aja-quality 1970s afro-rock?

 

Thanks for sharing.

Excited to try this on 64kbps punk and hardcore.

 

Last edited: Jun 23, 2025 at 8:20 AM

For those who are interested in, I've recently ported AudioDelossifyer (both code and models) to OpenVINO for local inferencing on supported Intel CPUs/GPUs/NPUs:

https://github.com/MarcoRavich/OV-AudioDelossifier#readme

Null test result:

 

AudioSR and Delossifier null test comparison:

 

I tried it on 1901 wax cylinder recording(40kbps/22kHz) and other (64kbps/22kHz) mp3
And it did not work at all, probably trained with higher bitrates

 

Interesting! Thanks for this "so called" infos

 

I started reading this post and I thought you were joking but then I realized you're not, are you! If it had worked I would definitely have wanted to hear what that sounded like.
Are we talking about the kind of stuff that Peter Jackson was using to restore dialogue for the Get Back documentary? I don't have a leg up on this kind of tech at all yet.

 

its not lossy

 

For those not interested in using gcollab or hugginface, MVSEP added the Apollo models a while back and recently the AudioSR and FlashSR models. They also have an explanation of what they do with demos and user generated demos on their news page here or the direct links Apollo with demos and AudioSR/FlashSR info.

@ArticStorm Had a similar thought to you, in that if it adds high end to compressed files let me try it on stems. I've run various types of separated stems through them and it varies depending on how useful the additional high end frequencies are, like additional transient info on drums and some vocal stems return added intelligibility.
To further your point maybe the generative audio models could be good at re-synthesising missing content from isolated stems if given context. Suno and Udio already allow you to remix or extend any uploaded audio, so maybe they could be tuned to focus on restoring missing info, just a thought.

 

Last edited: Jun 24, 2025 at 12:38 AM

...of course isn't, since is the difference between the Lossy and the Delossyfied signals (to better understand what it generates) !

 

Well, as the author itself claimed here "Better training data could lead to better results": you're free to train a model yourself (and, share please) !

https://github.com/kroll-software/AudioDelossifier#readme

 

...but MVSEP is (almost) the same of using gcolab or huggingface, just better interface.

Anyway:

...but they don't shares any (own/3rdparty) model, if I'm not wrong...

 

Last edited: Jun 26, 2025 at 8:53 AM

Dude, careful what you wish for, it could turn out your punk collection is actually a bunch of 80s hair metal and that Aus-Rotten album you've been listening to all these years is actually Mötley Crüe.

 

I came here to write, you don't need blind listening tests or whatever: just null tests. Your eyes are all you need.

That said, not to pour too much vinegar or acid on your parade, but I've seen excellent ai upscaling of 240p, 360p, and 480p videos last year. I think I saw a 480p file upscaled to 1080p and 4k with stunning resulting quality.

The MPEG I-layer II and layer III audio compression algorithms are nothing like the JPEG and MPEG I-layer II picture compression algorithms however. One uses psychoacoustics, masking frequencies and frequency discarding to render its files, the other uses other calculations. My point being, I'm not so sure an ai algorithm can restore audio spectrum that's been deleted altogether - whereas it's fairly easy to analyze a line (a p line) and double it, and render a seemingly higher resolution picture or video.

But I'm no longer in the game, so perhaps there's some new development in the industry I'm not privy to - good luck to you.