Hello Guys!

We all know, how the recent situation with mainstream music is: most releases by artists are mastered very "hot", meaning they are too loud. To achieve this, mastering engineers use brickwall like compression which often has the effect of introducing digital clipping. With these releases you listen to a wall of sound, often combined with digital distortions. This can cause fatigue (it certainly disturbes me) and you won´t have the pleasure of enjoying true dynamic. Read all about the Loudness War here.

Thankfully there exists a neat little piece of software: iZotope RX. This software aimed at audio restoration has the ability of restoring peaks in digital audio that were clipped during recording or mastering. Though this purpose mainly is intended only for simple correction I´ve found out that it works very well with CD releases that are victims of the loudness war. For sure I would like to know how it works, maybe someone of you has better means in discovering that. I guess the DeClipper in iZotope RX uses some kind of interpolation or extrapolation once it encounters digital clipping. From my experience it looks as if the processor "knows" how to reform the clipped material to its original state. This declipping for sure is not lossless and because of the impossibility to fully restore clipped material it can´t restore everything (according to my experience it restores about 60-70 %).

As it should be expected, this processor changes the sound. The change is not big but I often think that the declipped material sounds - apart from the erased distortions - a bit "muffled". Higher frequencies seem to be a little lower in volume. This may have its reason in the removed clipping (which usually sounds harsh). It even seems to have an effect at deeper frequencies: drum kicks have more punch. Also, the sound doesn´t seem so "stuffed" and "undifferenced". Only my opinion of course.

I´ll give two examples, the first is the song "Otherside" from the album Californication (Red Hot Chili Peppers), the second "Hollywood" from the album American Life (Madonna). The RHCP song is mastered approximately 12 dB too loud, the Madonna song 7 dB. Both songs contain high amounts of digital clipping. In both cases I used iZotopes RX DeClipper to remove distortions. As comparison, I produced another file that was just reduced in volume in the same amount as the declipped (RHCP: exactly 12 dB, Madonna: 7.3 dB). I have to lower the volume of the material so that it won´t clip again. That way the processor has more headroom for restored dynamic peaks.









I will offer these 30 second tracks for you to listen to. You´ll be able to compare the replaygained file to the processed file. You´ll be able to hear yourself. Here´s the link to the thread in the Upload section.

After you´d listen to the material I would like to hear some opinions. Do you think that all of this is pointless excercise? Or do you think, that this could be a valid and useful option to really improve the sound as long as we don´t have the possibility of getting the original master tapes? Furthermore, if these results prove to be good, could this DeClipper be done with open software? And before you ask: No, I´m not working for iZotope. I´m just a fan and I´d simply like to share my happyness with everyone else.

it is important to note that it doesn't improve the signal objectively, only subjectively. information is lost in the dynamic range compression and clipping that cannot ever be restored without access to the master. the filters in this program only further distort the sound.

This reminds me of an older thread: Partial Restoration of loud CDs.
There it started out quite promising, but soon turned out it was doing a good job on some samples and not so good on others.

Maybe. It doesn´t subjectively sound like there is any distortion left though - and no added distortion that is observable. But you´re right, I suspected this myself. Therefore I clipped several pieces myself on purpose with the leveler in WaveLab, only to be declipped by the DeClipper afterwards. And this processor came close, it restored about 60 % - 80 % of the dynamic peaks that these tracks had before I distorted them.

However, this DeClipper does NOT work well with every material. For example, if you set the Clipping threshold to low, then it will introduce Pops and Clicks. There is even a CD from Missy Elliott I can´t declip at all without introducing distortions. Another CD (Back to black from Amy Winehouse) also can´t be declipped at all. Some newer Loudness Maximizers obvisously can boost the volume without introducing digital clipping. And when there are no digital distortions, then the DeClipper can´t remove them. How did I forgot this? Sorry.

With the most material however it works wonderfully.

BTW, I did compare the frequency response of the two files with the RHCP song with another tool from iZotope: Ozone. This plug-in has a function in its EQ that can match the sound from one file onto another file. As a side effect, it shows differences in frequency response very good:


Compared to the clipped file, the declipped one has a different frequency response.
Less high frequencies - maybe because of the removed distortions?

Did you ABX? I admit I had some moderate difficulty ABXing the Otherside sample. For me, the changes were not only "not big" but for the most part not easily perceptible in a double-blind environment.

You are not right on this.
RX declipper only changes the waveform during clipped intervals, i.e. it does not distort parts of the original waveform. It only interpolates (or extrapolates) parts of the waveform that are lost. The interpolation method looks at the spectrum of the audio signal around the "gap" and synthesizes the signal in the gap having similar spectral content with the surrounding. If you listen to demos of the effect, the amount of distortion is significantly decreased in most cases, and it has nothing to do with low-pass filtering or something else that can degrade the whole signal. The only signal that is changed is the one that was already lost, specifically - the one that lies above the user-specified clipping threshold (lowering this control affects longer signal sections).

Cavaille, your sample of Madonna is not really clipped. It's processed with some "soft" saturation and maximizer. So, the declipper is not the "proper" way to restore this recording (although you may still like the effect).

The sample of RHCP is different: it's really heavily clipped and high-frequencies are overwhelmed with distortion. So, declipper is quite effective on it: the distortion is significantly reduced. The reduction of high frequencies that you observe happens due to elimination of distortion. It's better seen on a spectrogram display rather than on the time-averaged spectrum: vertical lines corresponding to clipping are gone, but useful HF signal features are intact.

(clipped spectrogram is at the top, declipped - at the bottom)

To conclude, I think that the use of declipper is justified whenever there is actually clipping in the waveform. However clipping is quite different from the brickwall limiting: although the waveform overview may look the same, zooming it in time will reveal presence or absence of clipping (histogram is useful too). Properly implemented brickwall limiting doesn't result in clipping.

Yes, I did ABX. With RHCP it was actually very easy. When ABXing I had only to listen for muffled sound and decreased distortions. The frequency response alone should be observable. With the track from Madonna it proved a bit different, there for me only the timing was a giveaway. Specifically, the timing of the Madonna track didn´t particularly improve. With the declipped material the bass seemed to "pop". Apart from that... "Hollywood" was difficult.

Did you program that processor also? Then I have to admit, that I´m a fan of your work. For proof of your sentence one has only to look at the two waveforms I combined with Photoshop. There you can see, that the DeClipper only affected the material within the -0.5 dB Threshold I configured. Starting with material below -0.6 dB nothings is affected as one can see in the fadeout of the song - which is exactly the same as the replaygained example. Thanks for clearing up, that this is not one of the usual dynamic expanders for clipped material. But how can this processor work with complex material and still produce these (to my ears) superior results?

That could explain, why the bass sounds a bit strange after declipping. But what does it "declip" then?

And what is this?

It is a track from Madonnas Hard Candy. Does this indicate that it is clipped twice?
If I try to set the threshold to incorporate both clippings I´ll get distortions like pops and clicks.

I don't think that this is 100% true. Yes, information is objectively lost, no question. But there is still contextual information in the signal that is not salvaged by a regular PCM decoder.

It would be right, if an ADC/DAC chain would output the following for clipped content:

unclipped signal - lost information

But it is actually outputting this:

unclipped signal - lost information + distortion

That is because the ADC uses the same symbol for the loudest possible volume as it does for louder than possible*. So a regular decoder just assumes loudest possible volume and produces the hard edged flat tops we know with a terrible kind of distortion. So always using a good declipper (as RX) could mean better PCM interpretation, that would be much closer to the original signal before ADC. Something like this:

unclipped signal - lost information + information reconstructed from context + slighter distortion

The only problem is discerning true clipping from hard limiting, since the bitstream doesn't tell us. But guys like Alexey, who have built such fabulous tools, could maybe raise automatic (true) clipping detection to a level one day, that would allow something like RX to be used in an always-on configuration.

*There even would have been a free symbol in PCM's asymetrical integer number range, but who could foresee that the music industry would broadly clip on purpose one day?

Does this mean the declipper actually uses the material that was distorted as a reference to re-build the peaks together with the process of looking around the clipped peaks for interpolation?

BTW, I did declip the whole CD - on my Pentium 4 3.8 it took some time... and now that the distortion is gone, I ask myself if the music was significantly altered with an EQ before boosting volume. Before declipping the frequency response sounded relatively balanced - now it sounds as if someone used low-pass-filtering. As if they knew how it would sound after maximizing the volume and in order to counterbalance that they changed the frequency response. If so, who the hell did master or mix it that way? That person should be shot. Another question... is the near monaural sound intentional?

I don't know what you mean by "material". Samples can be clipped. For example, imagine a case where you have 10 unclipped samples describing a steep rising slope, then 3 clipped samples at the top, then 10 unclipped samples describing a falling slope. If you interpret the three clipped samples as a flat plateau as a plain PCM decoder does, the sharp edges between both rising and falling slope and the top would introduce high frequency distortion. If you interpolate a nice, soft, and rounded 3 sample hill top from the 10 leading and trailing samples instead, there is no clipping induced HF distortion.


That's probably because the mastering genius attenuated the higher end because the mix had too much HF energy due to clipping distortion. If you remove that distortion, you get more or less the nature of the original signal without the distortion, but possibly still with the low-pass to fight this distortion in place.

But this must not be the case. It can already sound low-passed in comparison, because all that HF distortion, that you might already be used to for that record, is missing without an actual low-pass in place.

I usually do not listen to such music. I only bought the CD yesterday for little money because I wanted to know what people were talking about. According to Wikipedia it serves as a good example of Loudness War together with the album Stadium Arcadium. As this kind of music is new to me I assume I see it with a "fresh" distance. Of course, I compared the balance to my usual music - which isn´t remotely the same. I simply can´t apply the same preferences I have for classical music or orchestral towards rock music. Still, I miss the crisp sound I´m accustomed to and still Anthony Kiedis voice sounds as if he did have a severe cold during recording. This CD is remarkably unbalanced. There are huge frequency holes all over... I tried a little "re-master" - now it sounds not so bad after all.

It replaces the signal sections that are totally lost (flat lines) with sections that are reasonably well interpolated from the undamaged surrounding signal. The interpolation algorithms are rather complex, they can be technically described as expectation maximization (EM) over bidirectional least-squares autoregressive (LSAR) interpolation. In simpler words, they are based on LPC (harmonic) modeling of the signal and ensure that the interpolated section joins the surrounding data smoothly.


The algorithm doesn't care what lies above the clipping threshold parameter. It just synthesizes a replacement signal and puts it there. If the signal there was good enough (not too distorted), then the replacement can sound worse than the original.


This may indicate that clipping (or saturation) has been asymmetrical for positive and negative voltage. RX is not quite suitable for repairing such damage, but if my guess is right, you can try to symmetrize the clipping points by adding a sufficient DC offset to the recording.


This is a good idea. However the detection can become more difficult when the clipped sections are non-flat (e.g. contain some ripple due to filtering, digital/analog transfers or whatever) or when the signal had been subject to saturation (as in Madonna sample).

P.S. Thanks everyone for your kind words on RX, we keep working on it.

If I understand this correct, it guesses. Well, if so, then it is extremely good at guessing.

Wow, thanks very much, I´ll try that - and post the results afterwards.

I second that. If it could be used automatically, it would spare me a lot of work.

In my mind you should get the nobel prize or something like that. I use your resampler and declipper regularly - just wonderful tools. Expensive, but worth every penny.

I don´t know why I now can´t reproduce the results for this. Almost a year ago with a different rip done on a different drive under a different operating system with the same CD it looked like this:


But now it looks different. Maybe I selected more material today for the waveform window inside the DeClipper - but if I select less material that can be computed, it still looks different.


I´m awfully sorry, I can´t reproduce the results. What puzzles me the most is that I used different replaygain settings - for the older rip I used -8.7 dB, the replaygain scan I did today produced -9.4 dB. Maybe this is because of a different foobar2000 version.

Anyway, this shouldn´t puzzle any of you. Maybe I made a mistake with the earlier rip... don´t know, too much time gone.

Cavaille's technique would be quite achievable if a lookup table-based scheme for brickwall limiting were somehow standardized; her problem is that she's making assumptions about the masters that can't accurately be made, but if they were actually true, it would work pretty well. ie, something like a modified μ-law? Map the top 1024 sample values to an exponential scale so that -0.25dbFS maps to -0.25dbFS but 0dbFS maps to +6dbFS?

Anybody? Crickets?

It resembles HDCD

Very interesting. But so many CDs are mastered different... impossible to know them all, I think. But maybe there is a way - I´ll never give up hope.

Of course, the mastering engineer is trying to adjust the overall sound. So as compression/limiting/clipping affect the frequency balance they will make EQ adjustments for what he/she feels is the best (or most commercial) overall sound. And, from what I've read, it's common for recording/mixing engineers to make EQ adjustments when they use compression. (Recording/mixing engineers are usually more gentle with compression, and hopefully they are avoiding clipping!)

That could be caused by clipping/compression too. Let's say the mixing engineer carefully panned an instrument somewhat to the right... Let's say it peaks at 0dB in the right channel and -6dB in the left. The mastering engineer boosts/compresses/clips everything, and now the signal hits 0dB in both channels. (Anything that's centered, panned hard-right, or hard-left won't be affected.)

Sometimes a simple subsonic high-pass filter can "even-out" an asymetrical waveform (without adding DC offset). And, adding since DC offset can have undesired effects (like a "click" at the begining & end), so if you add DC offset, it would probably be a good idea to use a high-pass filter (to kill the offset) after declipping.

I agree with you completely. I´ve just done some reading - it appears that a guy called Vlado Meller is responsible for the mastering. But the true reason behind all this seems to be the producer Rick Rubin because he wanted it so loud. Rick Rubin also is the guy who produced Death Magnetic and Stadium Arcadium - it figures. Apart from that, shame on this Vlado guy. Over here at Gearslutz.com you can read what a "cool" & "awsome" guy he is. In that thread they basically try to compress and distort some example tracks to hell after following his example - and they love it!

Look here:



Posts like the following are almost completely ignored:



Which is a clear sign to me that some mastering/mixing engineers don´t have an idea when they are talking about good sound. Ah well... there I´m going at it again... my own taste. But if someone produces a record that is intentionally treated to sound like something from the engineering stoneage in order for that record to appeal to the masses or some idiotic, unbalanced sound preference... well, I just can´t stand it.

You know what? Tomorrow I´ll upload some additonal samples. Californication after a little quick "remastering" and filling the frequency holes according to my taste and the rule of balanced sound together with configurations so that everyone can do it themselves. Now that I have those restored peaks back I can change the sound to my liking.

OT: For Californication, there is an unmastered bootleg circulating around. Just Google it.

You might want to do a comparison with it, but IMHO, it sounds a bit "muffled".

Thank you very much for this information... but I´m not very much interested in bootlegs. You never know about the quality and I like to take care of things myself.

And as an example I offer you three sample of how I´d do a master for Californication. I admit, it sounds a bit too "beautified" but at least it´s more pleasant and Anthony Kiedis almost sounds as if he hadn´t had a bad cold during recording.

Try it out for yourself.