Stereo Tool

Hans,
Did you tryed my, maybe stupid, but simple idea for Clip level detecting. I wrote it somewhere already.
Simple Peek detection with double fast AGC on input and output.
- 1st Peek detection Level finds largest peek and adjust, and 2nd just folows 1st and do opposite to maintain always same level.
- remove some highs from input detection to avoid peeks from highs.

Bojcha, I don't understand what you're saying at all... But besides that: Clipping level detection alone doesn't suffice! I really need to know PER SAMPLE whether it's clipped or not. Because for tracks like the Metallica song, if I use the clipping level I need to throw away more than half of the samples, which makes it impossible to properly reconstruct the audio. And many of the samples I would be throwing away are actually ok........


Edit: At the NAB in Vegas, we had to reduce the clipping level determined by the Omnia 9 by 2.5 dB to capture all the clipping from this Metallica track. That way it still sounded MUCH better than without declipping - but 2.5 dB means that we were REALLY throwing a way a lot of useful data.

(What apparently happened in this track is that after clipping, something like equalization was done). There are more tracks like this - although for MOST tracks clipping level detection is extremely easy.

Don't you want just to have "Assume values to be clipped if they are at or above" slider to be automatic ?
or something more is needed ?

Bojcha, I just edited my post above - yes, I need more.......

ahh ok then ...
can't help further

I notice in the Metallica track that the clipping happens at different levels.
Idea: make something which detects the flat lines with a maximum difference per sample too allow some wobble in the flat line.

@ eldoradofm: That's already part of what I'm doing. But it's too simple because:
1. The line is NOT really flat - there's actually quite a lot of 'wobble' in it. And (wildly varying!) tilt too.
2. Clipping can also be very short, if that happens there won't even be a flat line visible in the waveform...
So I'm actually doing something different.

Unfortunately, I'm getting too many false positives: If you play for example a bass sound, the tops of the bass will also look like long flat lines - and my current algorithm unfortunately detects them as clipped.

Could you post images of the different clipped samples it has to detect?

Ah! I just found at least part of the cause....


Edit: Ok, I've been looking into this a bit more and I need to do some more thinking first. I think I'm pretty close to the solution (the Metallica track sounds really good!), but I need to fix some remaining issues.

Flat area detection might work, but the real problem is that I have a filter that determines how much clipping occurs, and that's the filter that gets confused by loud sinusoids in the signal. Next, I use the determined amount of clipping to mark samples as clipped based on the likelyhood that a specific sample is clipped - the problem is that if the number of samples that's being marked as clipped is far too high, this results in throwing away far too much valid data.

--> I did just find 2 bugs in the algorithm that determines clipping likelyhood...

So I first need to fix my filter that determines the amount of clipping, after that the rest should be easy (because most of it is already working).

For the Metallica song the current filter works, and for extremely clipped audio it also works - the biggest problem is currently that loud sinusoids (specifically loud bass) are considered to be clipped while they are not.


NOTE TO SELF: First add clipping distortion level detection. - will continue later...

That's difficult - because in some cases it's not visible at all in the waveform that clipping has occurred.

You've seen the Metallica track, that's one extreme example.

No time now to create some screenshots, but imagine that this audio is clipped: http://www.audiomisc.co.uk/asymmetry/miles8.gif

How would you see that? Especially if only a few of the spikes are clipped, and if the clipping level isn't completely constant... (I'm not saying that my filter will be able to... But I hope it will).