DSP Robotics Support

Hello!

The last weeks i tried a project that should restore high frequencies to audio files that are not high-res or just saved as mp3 recording. I got to a result now after some days of work and wanted to ask for feedback.
I am planning a release of this tool but not today, but I will show you a converted file.

Preview:


As you know restoring something that is not present anymore is quite impossible - it is that impossible that it took quite long to get an idea how to restore something. So, if you restore something it is not the original - the goal just was to get a better sound - better than low quality mp3 and also better if you do an up-sampling from 44,1kHz to 96kHz.
Somebody in the old SynthMaker forum said, you can change a cow into a burger but not a burger into a cow.
I may have to add something here: You still can make the burger look like a cow.

You can look down for some theories / ideas I did here.

I need feedback:
I would like to know if you like the result and if you may have additional ideas here.

I have taken an “open source track” I linke from “TheFatRat” – “Monody (feat. Laura Brehm)”
YouTube: https://www.youtube.com/watch?v=B7xai5u_tnk

I got this music as lossless wav and converted it to mp3 128 kbps (to here more restoring)
I also made an experiment to upsample it from 44100Hz/16bit to 96kHz/24bit

I made Multi-Track-MKV/MKA files – If you play the files with VLC-Media player you can go to “Audio -> Audio Track” and select the Audio Track you want.
Download Multi-Track-Audio: https://www.dropbox.com/s/acx0nsenjno48 ... D.zip?dl=0


Well if they all sound the same, you may take a look here: I did speed the track down to the half speed. (Here you will here the restoring operations):
Download Half-Speed Multi-Track-Audio: https://www.dropbox.com/s/i4uqwo26twqv5 ... D.zip?dl=0



Some theory behind this:
My Denon AVR (and also Maranz) has a restorer for lossy audio materials. As I wathed the spectrum what this thing does I found out, that they add high frequencies.
You all may have herd from harmonics that an additive OSC will generate. Those harmonics are calculated by a row series. So, the harmonics follow the same scheme each octave falling off a constant value in db. If you need more high frequencies you just generate more through the “Fourier series”.
Music Instruments operate most from 20Hz to 4200 Hz. Every frequency generated after 4200 Hz are harmonics and follow the scheme of the falling Fourier series.
So why not taking the last frequency were a song has the "fall off" because of samplerate, grabbing the last octave with a sinc-Kaiser filter and pitch it up with a phase-neutral pitch-shifter and applying some negative gain-offset. The result was that great that I got interested in this.

My words to Mp3:
Well, LAME mp3 for example uses this Fourier series to get down the data-rate. But a Mp3 file with 128kHz will lose frequencies especially all over 15-16kHz. I do not get more fare into this – if you really like listening to music you should not have mp3 under 320 kbps…

CD / Wave / Flac:
A CD-Audio with 44100 Hz would be able to store frequencies up to 22050 Hz. Most people hear till 18-20 kHz. So, you might think that this is enough. But listening to High-Res audio still sounds more interesting than CD-Audio (Or mp3…)

But getting real High-Res audio is quite difficult – I self had to complain at a high-res store that offered just simple upsampled 96kHz/24bit that came from the original CD with 44100 Hz.

And sometimes you do not get any better quality (Live-Concerts, low quality mastering, …) – so you need a restorer.

And yes it "works" and makes a little enhancement.

MP3 Spectrum normal speed:


44.1 kHz to 96 kHz half speed:


Kind Regards,
C.Hackl

Edit: corr. links.

Edit: Link for alpha release post:
viewtopic.php?f=3&t=36500

I get 403 Forbidden for those links.

I once made an enhancer which worked on a simple principle. Take a HPF filtered signal, add some distortion and subtract the filtered signal from the distorted one, so you’re just left with the HF component from the distortion. Then add that to the whole original signal. It seems counter-intuitive but I thought it worked quite well. Not my idea though; I read about it somewhere.

I’d like to hear your approach.

Cheers

Spogg

Same here - 403 Error for the links.

I like your pragmatic attitude to the problem, though! It makes sense to me that using the "Fourier Series" rather than distortion might be a more pleasing way to do it for full mixes, as distortion is likely to generate aliased frequencies unrelated to the original harmonics. If you can get those links fixed, I'd be interested to hear it.

Well i have corrected those links to dropbox. I hope they are working now.

Using a distorsion may rely be an option - and i also found plugins that are doing enhancments by distorsion like this one:
https://vstplug-ins.com/soundrecovery.html

But i was not that happy with the result in hi-fi.

In my opinion distorsion is a good alternative if you have low cpu (Media Players and rly the most have mp3 enhancers). Converting the Pitch-Shift is taking fare more cpu

Pls let me know if the Links are now working.

Yes I got the files!

I compared the restored one with the original and to my poor old Vulcan ears they sounded pretty much identical.

When I compared the MP3 128k with the original and restored one, I could certainly hear a good difference.

So I would say this is a good system you made.

As a side issue, I didn’t know you could encode multiple audio tracks like that. I tried to extract the poor quality MP3 to compare with my old distortion-based system but I couldn’t extract it. If you have a minute could you link to just the MP3 and I can maybe do a comparison?

Cheers

Spogg

Hi!

Yes this would be great if you can do this - If you send me the restored file in 44100/16bit flac i could make a half-speed version again, to hear the exact difference.

Mp3 File:
https://www.dropbox.com/s/qnqf970nwis24 ... D.mp3?dl=0

flac File (Original was wav and i converted it to save some disk-space):
https://www.dropbox.com/s/ru8l376cpdulh ... .flac?dl=0

To your side issue:
I used mkvToolNix to piut the tracks together - as it is much more comfortable to compare.
You can split them again if you use "gMKVExtractGUI":
https://sourceforge.net/projects/gmkvextractgui/
But you need MKVToolNix:
https://www.fosshub.com/MKVToolNix.html

So just in case if anyone wants to extract the audio tracks.

Regards!

Interesting work. I would assume that above 10 kHz the human ear is not too picky about the spectral or harmonic details. White noise with an envelope derived from the upper range of the lossy material cold help already. It might depend on the source material, though. I am referring in particular to human voice recordings.

Very impressive stuff!
Can't wait to testdrive the tool. And it would be just great to be able to do batch file conversions too.

By the way, I dig that you 'start conversation' with us at the FS forum, but with that button in the GUI I assume you mean 'start conversion'.

Yep this seems more suitable
I'll fix that - my english is a little poor.

A batch conversation was planed - but i did get away from it because each track need some spetial optimations. Since each track is mastered a little different and has an other harmonic falling scheme depending on the sounds of the instruments used.

I used a 12dB/octave HPF set at 7.5kHz overdriving a Tan(h) to create HF distortion:

https://www.dropbox.com/s/bhu9lvh15s05w ... D.wav?dl=0

I used the actual MP3 file in Reaper and exported the output recorded from the enhancer.

I think I added a bit too much of the distorted signal because it sounds a bit brighter than the original Flac file.

One thing I noticed was the sound was nice on speakers but on headphones the stereo positioning was not so well defined. I suspect this is due to the phase shift in the HPF filters blurring the position cues somewhat.

In conclusion, I think your system gives better results.

Cheers

Spogg