seeking knowledge

[Galliyummy]

I'm very interested in this topic because I am a DJ as well as a programmer. I stumbled across this repo after some meticulous digging into AI audio processing and upscaling. This issue is to seek advice for me and anybody else who may have a question like this. I have an old mp3 at 320kbps and I am looking to use AI and neural networks to upscale it. I started by trying to use fatllama but got hung up on the final audio processing step (will do more testing later). Then I stumbled across DSRE and tried using that. After some translating, I was able to make a solid gain in increasing the quality of my file. If anybody has a better direction to point me in towards this field and getting my compressed file to be closer to lossless, any help would be greatly appreciated.

[Galliyummy]

Also DSRE cut off the time of my mp3, so I might have to make slight adjustments on that front.

[MarcoRavich]

Hi there, I'm the author behind HyMPS project.

My suggestion is to wait for the next release of Intel's OpenVINO™ AI Plugins for Audacity - scheduled for the first week of October - since they'll support some cool restoration models such as Apollo/Universal:

Check out this interesting discussion about delossyfiers @ their git where we shared some cool resources.

I also suggest you to try HRAudioWizard - a non-deeplearning audio enhancer which outputs very interesting quality results - on your file.

Last but not least, don't miss the cool @deton24's Make your own remaster paper.

[Galliyummy]

I'll definitely be on a lookout for that Audacity plugin but in the meantime I can show you some results that I put together:

This is the song that I am trying to restore and the one that inspired me to look into this in the first place - J.A.W.S by Lxury. As a DJ, I have been trying to get ahold of a higher quality version for a while, and I never really could.

This is after I used fatllama to get a flac file:

This is after I converted it into the format that I want using my own foobar2000 preset:

This is after I took the flac that I got out of fatllama and converted it using the default settings, as I ascertained that they would be proficient for the results that I wanted to get out of it:

This is after I put that same result using the same foobar2000 preset, this is also my final result:

[Galliyummy]

Also interesting to note that DSRE seemed to have added some high frequencies that are not just enhancing the hats but they seem to almost mirror the melody.

[Galliyummy]

HRAudioWizard seems to have difficulty saving the output flac, generally it gets corrupted, so I will look into that more later.

[Galliyummy]

Starting to use openvino, I think that this might give me my best result so far, will test and see. the openvino plugin has one issue that I need to address, which is that it can only run on my cpu or my integrated gpu. My dedicated gpu does not work, as openvino is not cuda based. For me to attempt to restore one singular channel of audio takes an estimated hour and a half.

[Galliyummy]

After doing yet another round of research, I have found that openvino's super resolution is ported from this repo: https://github.com/haoheliu/versatile_audio_super_resolution. I am going to clone that and see if I can clone that and mess around with it.

[MarcoRavich]

Hi there, I've performed some non-deep learning audio enhancing algorithms tests too and here are my results...

Source MP3:

Fat llama + DSRE enhanced:

HRAudioWizard enhanced:

note: I've previously decoded the MP3 to WAV using foobar2000 to avoid any possible decoding flaws

Last but not least, I invite you to join the AudioLDM discord server or Hydrogenaudio's 3ad to talk about your tests/results.

[eevan78]

Hello,
Just one remark. As lossy compression is heavily utilizing psychoacoustics and frequency masking effect, it is almost useless to judge audio quality by looking at windowed frequency spectrum plot.

What I can see from these graphs that you have attached, especially on the second one, you have images of the original spectra around 44,1 kHz and 88,2 kHz. Then it appears to me after comparing final step with previous that there were some aliasing from those images into the baseband. Which is always undesirable.
So I'm not sure what is the benefit of the complete process...

Original source that is MP3 with bitrate of 320 kbps should be audibly transparent in vast number of cases. Unless some bad encoder was used, or music has some specific passages difficult to encode, there should not be any issues with it.

[MarcoRavich]

Yes, "restoring" 320Kbps MP3 is quite useless expecially using non-deeplearning algos.

As well explained by @Sg4Dylan - the EmiyaEngine's author - in his repo:
As long as there is the will to become the real thing, the fake thing will be more real than the real thing.