Stereo Tool

Michi: The two have identical volume for most of the audio, except for the added spikes. So there's no need to change the volume.

I have to admit that my first impression was that there WAS an improvement.


Bojcha: I just downloaded and tried SeeDeClip on the sample I posted earlier. Either I don't understand how to use it, or it doesn't work at all - but it sounds horrible! Have you had any more luck with it by now?

From my point of view...

This feature, in Omnia.9 is for restoring clipped sound form mastered CD, which is case nowdays.
But actually point is that in same time that same cliped sounds, later in processing sounds better, meaning Punchier, warmer ..
If we assume that already Mastered loud track are more cliped and automatically less or no punch, this filter from Omnia.9 will make it more dynamic and as much as possible get closer to not-loud and already dynamic tracks.

This is also realted with new MPX cliper. We can clearly see almost 3dB louder de-modulated Audio level at output, which actually gives room to less use limiter (loudness) and inn same time get on Dynamics!

All is done do keep dynamic and in same time loudness, loud tracks to sound as much as possible dynamic too.

SeeDeClip,
Here is the Lady Gaga - Alejandro, one drum constantly clipped


And here is SeeDeClip


Listen to both .. i can't hear some spectacular difference.
So i played both thru ST.... Again.. no difference!

So, this what you make Hans is dfferent. And for me much more useful !

EDIT:
I just Player both tracks through Breakway .. again do big difference
or maybe i just used bad example

I fed this track (Lady Gaga) through the new Stereo Tool filter, but it didn't have much effect either. Which makes sense: It is not really (hard) clipped, but filtered too (the clipped parts are not a straight line but still contain some other frequencies). Ie. it is processed by something similar to Stereo Tool's Loudness filter.

I think it's probably possible to correct this as well, but it's a bit more complex (still thinking...).


The filter I have now only restores real clipping. I just tried it on 'Red Hot Chilly Peppers - Californication' and on that it does have a much bigger effect (peaks to over 200% get restored). Feeding the result through a tamed down version of your preset (lower AGC output setting, lower Loudness setting) results in a warmer sound.


By the way: I found some good news: I've tried what happens if I try to restore the audio of my clipping example after first feeding it through an MP3 compressor. And it still works! (Although I need to ignore some more data, which _could_ potentially even _cause_ distortion...).

Only for some reason almost all the MP3s I find online suffer from (sometimes very extreme) tilt. Often I cannot even SEE any distortion if I look at the waveform.


I'm currently trying to find a way to 'keep' all the information that is still contained in the (not completely flat) samples that I'm throwing away. If I can manage to do something with those I can be more sloppy with determining which samples are clipped and which are not - that might greatly improve the performance for 'almost' clipped audio such as in this Lady Gaga track. And at the very least it will improve the sound for MP3s, where the line is never completely flat. (Basically I need the algorithm to 'estimate' the correct value for clipped samples, but keep the original if there's no disadvantage to that).

(By considering too many samples to be 'distorted' I'm in fact adding distortion... Which I then need to clean up.)


Basically I need to make sure that my algorithm chooses the original signal if it doesn't seem to be much worse than the reconstructed signal. And if possible it should do that in every iteration of my filtering process.

Because I'm a real artist, here's a drawing of my problem:


Say this is either an MP3, or a recording clipped with filtering (such as the Lady Gaga track).

See image 1. To be able to restore anything, I *need* to consider the samples above and below the red lines as clipped.

See image 2. Hopefully, restoration leads to this, which is very close to the original, except that the bass has been un-clipped.

See image 3. But it could potentially also look like this.

The closer I move my red lines to the top, the bigger the chance that I will fail to recover the bass. But the further I move them away from the top, the bigger the chance that I fail to recover the highs.

Ow, and to make things more difficult: The sharp angle in the "clipped" bass waveform also contains high frequencies. So I cannot separate them.


I do have a possible solution, but it will cause the CPU load to get insanely high. Currently processing 1 minute of audio already takes 15 seconds (SeeDeClipper uses a few tens of a second, but the result is also much worse).

huh...
question: "clipped" contains highs but when it's recovered, it should contain less highs, or only harmonic form that high tone will be at output ?

Well, that's the problem. Recovering means complete removal of those highs. But in the case of image 3 also a sound that should not have been removed is gone. So both are treated the same way - but they should not be. And there's not really a way to separate the two.

If highs are also clipped, they should come out LOUDER instead of softer! But other highs (clipping artifacts) should be removed....

Also remember that I draw an extremely simple example here: If a few more frequencies and a hiss sound are present it looks like a complete mess. And all of those sounds could be clipped and have harmonics (that's what makes the examples in the .WAV file I posted earlier so difficult).

I don't want to disturb the progress of your declipping discussion, but I have to relativise some things in conjunction with this Undo magic.WAV.

Maybe I had one or two beers too much Yesterday.

Today I am not sure what version I prefer.
Though I generally prefer loud drum sounds, now I hear that this UNDO processing causes damages to the voice of the singer (instability).
It sounds a little bit like standing close to a german autobahn (freeway) while cars are rapidly passing from left to right and vice versa.
Yes, this is annoying.
But as I already have stated earlier that this is based on too much expansion (less would be better here).

Another example: Rihanna - What's my name

Michi: On that other forum you requested a 60 second restored version of American Idiot.

This is what Stereo Tool makes of it.
http://www63.zippyshare.com/v/69386596/file.html

Unfortunately, due to the fact that I haven't fixed the issue I described above yet, I had to clip another 11% before I ran Stereo Tool over it... Still the dynamics have improved (it's not very noticeable, but the waveform clearly shows much bigger spikes at bass kicks).


EDIT: On that other forum, they have now also placed a reworked version... But they didn't use an MP3. So I searched a bit further and found a file where I don't need to clip 11% extra myself.

Here's that version, set at the same output level as on the other site for easy comparison.
http://www63.zippyshare.com/v/54616330/file.html

This version sounds much better than the previous one I posted (apparently even 11% extra clipping wasn't enough there).

Ah! New idea that might work to avoid damaging audio: Use 2 steps. First, step 1, treat a lot of samples as if they are clipped. Then filter out those that after reconstruction are close to the original values, and continue with the rest.

Step 1 can be relatively fast because only a coarse estimation is needed.

This may leave some soft clipping unfixed, but I only need to do this for samples which I'm doubting (samples that I *know* to be clipped should never be filtered out).