I have to admit I've been wrong about something. See, many believe that true peak limiting sounds worse, but the matter of the fact is we've been using it wrong.

And on my last video, I had multiple people ask if I could debunk or look into true peak limiters. I had a theory that I'd present mastering engineers I knew a bunch of blind A/B tests, I'd plot the results, and then I'd call it a day.

But that didn't happen because I reached out to Fabian from TDR Labs, and he's an incredible developer and has a great knowledge of DSP.

He explained to me everything he knew about true peak limiting, and that put me down a rabbit hole—a rabbit hole of research, learning, sinc wave functions, Whittaker-Shannon interpolation formula, having a friend compile some code for some tests I needed to do, get a confirmation from iZotope's engineering team on their true peak limiter, and putting this video together.

So stick with me; this is a meaty video filled with my research, interpretation, tests, and recommendations.

I invite anyone that has information on the topic to share, so in the comments, challenge my ideas and we can learn and be better engineers together.

These are videos which I get excited to make, and if you get excited to watch them, don't forget to smash the like button and subscribe for more.

So, all limiters measure their max peak and reduce gain relative to it; the difference is in how various limiters interpret that signal.

Interpreting the signal—now, this is what differentiates one limiter from the next. An analog limiter measures the level of a signal, which is a continuous waveform, so it's always got a true reading on it, whereas digital limiters measure the level of a signal from the PCM encoded Pulse-code modulation is the method used to digitally represent sampled analog signals.

Now, you might be familiar with looking at something like this.

However, each of these samples doesn't represent an instant moment in time. In fact, each of these samples is a scaled sinc wave, which extends, in theory, infinitely wide, both past and present.

The resulting waveform we see in our DAW from a stream of PCM samples is a sum of these scaled sinc waves, and the sum of the sinc wave values represented as a waveform, which we digitally see in our DAW, can produce over- and undershoot below and above the given peak of the given sample value.

Now, this presents a significant issue with the accuracy of digital limiters. At most conventional sampling rates, 44.1 or 48k, there can be up to three decibels of difference between the discrete PCM sample values and the resulting waveform, which means there's often audio passing through the threshold of a digital limiter without a corresponding PCM sample point for it to react to.

Now, this is an issue and, in fact, a major flaw in the function of normal digital limiters. So, how does a true peak limiter seek to solve this?

To achieve the same physical behavior as analog, true peak digital limiters emulate the D-to-A converter and measure its output digitally.

Speaking one way to emulate the D-to-A conversion process is to interpolate more discrete values in between the present ones, so the limiter can measure the level of a signal from a more accurate sample point.

ITU-R BS.1770-4 lays out a consideration for accurate peak metering of digital signals.

Now, with the standard of four times oversampling, this minimizes a sample peak underread from three decibels down to 0.668 decibels, and at eight times oversampling; this minimizes a sample peak's underread down to 0.169, which is the same standard used for FabFilter Pro-L 2's true peak limiter.

I found this article published in 2015 by iZotope's principal software engineer, where he lays out some code in Python to compile this sync wave offset a fraction of a sample.

The sample point has been offset by 3/8 of a sample, or 0.375. Now, I actually failed to compile this code, so I reached out to a few friends within the programming space, and one of them was able to compile the code and provide me with the necessary PCM-encoded sync wave for my tests, which is this one here.

So, a big shout-out to Kevin, thanks for that.

The test is simple: given that we've now established that limiters function by how they interpret the signal, a digital limiter interprets that signal by the PCM-encoded sample peaks.

PCM-encoded data can underread up to three decibels at 48 kilohertz. By interpolating new sample values, we minimize this underread within 0.668 decibels and 0.169 decibels using four and eight times oversampling, respectively, as calculated and presented in the ITU-R documentation.

So, can I get FabFilter Pro-L 2 true peak limiter to null with FabFilter Pro-L 2 using eight times oversampling when setting the threshold to the PCM-encoded sample peak?

And in a second test, matching the threshold to the reconstructed or interpolated true peak.

Now, with this shifted sync wave, we have a sample point at negative 8.13. So here I have a true peak limiter set to 8.13, and here I have a normal limiter set to 8.13, FabFilter.

Pro-L 2 interpolates the data at eight times oversampling, this one does not, so what we're going to do is we're going to play this back in a loop and then what I'm going to do is in the middle of that loop I'm going to switch eight times oversampling on here and null it, and you'll see that it almost nulls absolutely perfectly, well below the noise floor, so let's do that now.

Okay, so one thing we notice is the True Peak limiter is definitely doing some limiting; the normal one isn't recognizing that sample point. But what happens when we turn oversampling on?

Ah, we've reconstructed sample points, or interpolated sample points in between the existing ones, and it's now recognizing that it has to be limiting.

It's actually missing that information. So take a look at this null test, when I—so it's about negative 140 decibels is the difference, indistinguishable to the naked ear.

If I turn this off, look how much higher it is, up at negative 80, negative 75; it's absolutely doing—it's actually doing nothing there.

So again, I'll put the oversampling on, it completely nulls. So that's me doing a test there between the oversampling here and eight times—I mean, the True Peak limiting here and oversampling here.

So what happens if we set this to 6.37, okay, and then we're going to interpolate this data by upsampling it, because what's actually going to happen is the other limiter won't pick up anything at all.

So if we resample this, we're going to have new data points up here, and one of those data points just happens to be at 6.1, which is perfect because it's just right over that little threshold here, the way it's interpolated.

Now, look at this. So this one here is set to 8.6, sorry, set to 6.37.

True Peak there, and this one here, we'll bypass that. We'll turn the oversampling off so we can hear what a normal one and a True Peak one is doing, not here, but analyze what it's doing, and then we can put the oversampling on, and we'll get a new data point to work from.

So here again, because there's much less gain reduction, they almost null anyway. But the whole idea of this test and exercise is to show that the oversampling then allows this limiter to start working and recognizing those sample points above it.

The null becomes much more, like it's almost nothing there.

Um, so yeah. So what does this mean in practice? Well, True Peak limiting doesn't have a sound. The gain reduction algorithms, look-ahead, attack, release, and ceiling all theoretically should function in the same manner.

Simply put, it measures a signal with greater accuracy, and when that accuracy reduces underreads from three decibels to 0.668 decibels at four times oversampling, that can create a considerable amount more of gain reduction that wasn't present in the signal before.

Now, if you choose to turn True Peak limiting on but want the same amount of gain reduction, dial back the gain or threshold of your limiter around two to two and a half decibels.

Then you'll have a similar ballpark amount of gain reduction as the native PCM encoder limiting. But the best bet is to throw the voodoo of True Peak limiting out the window and just start by oversampling your limiter at four or eight times, not only because it minimizes underreads and overshoots but it helps the limiter function with greater accuracy and reduced aliasing distortion.

Consider especially the aliasing distortion. Now, everything I've spoken about is pretty fresh information to me, and I'm going to do my best to answer all the comments as best as I can in the comment section below, so please feel free to share with me your experiences with True Peak limiters and any knowledge you have on the technology below.

Is True Peak Limiting a Lie? What You Need to Know!

Hello. I'm Nicholas Di Lorenzo. I'm a mixing and mastering engineer and owner of Panorama Mastering Studio in Melbourne, Australia. I’m an Italian-Australian. born and raised in Melbourne.

This channel is for anyone involved in the music making process, artists, engineers, producers or musicians.

To stay up to date, hit the "bell" icon for alerts as I release videos weekly on all things audio, mixing, mastering and industry.

-

Want to learn more about what I do?
https://panoramamastering.com.au 
@Panorama_Mastering

Is True Peak Limiting a Lie? What You Need to Know!

Scripts