If you say so...

Alistair

Remenber, It's TV. You can mix it like a "feature" but somewhere down the line you have to hit specs or a brickwall limiter. Better you(the Mixers) do it than
someone else. Lower your threashold of your peak limiter.
You said you were off by .9 of a DB. Don't redo everything just for that.
Like TPAD said adjust , peak limit move on. You have taken far too long to post this. Listen, fix and go! DONE!!!

That's only for one example. As Alistair correctly pointed out, it can be as much as +6 over 0 dBFS or even more. This is a repeatable phenomenon that has been examined in several peer-reviewed papers. It's been discussed for years; it's just that with recent technology and broadcast's clamping down on specs, we're faced with taking the extra step to make sure it clears. Dan Lavry's paper is excellent and his converters are some of the best on the market. This argument is not new, btw, it's been beaten beyond death in mastering forums years ago.

Hey Alexey,

Have you checked this out?

Solid State Logic | Music

I know it's AU, but you could use something cheap on the Mac that works with AU just as a final QC.

Then why is it KK that I have never had a QC report over spec?
They are mixing TV but they think they are mixing a feature.
It's TV specs. They are also missing it by 9/10th of a DB not
6 db.
And don't use the L1 if you want something to sound good. It doesn't.

1. You've been lucky and haven't had program material that generated it...yet. Are you saying that this doesn't exist? Lots of mastering engineers
(including me) would disagree. As stated earlier, the Pohlmann text is illustrative, although it isn't elementary math.
2. TV is more restrictive than theater because they have to comply with FCC and now ATSC specs.
3. .9 dB still trips the meter.
4. When I was a sysop, I spent several hours on CompuServe chatting with Michael Gerzon whose groundbreaking research led to L1 among other Waves plugs. It's definitely an early implementation which has been superseded by later variations, but as with anything, there are ways to use it and not to use it, although I haven't used it lately. Still Gerzon was an excellent mathematician and it is he who first noticed the phenomenon, if I recall correctly. According to him the theoretical max possible clip level was +15, far above .9!!!! Of course you can brickwall limit a couple of dB below threshold, but then you're looking at -8 rather than -6 (assuming -6 is the spec) and you run into critical people who listen with their eyes. If you aren't hitting -6 they think you didn't do it right. Better to have something that will trigger the clip and let you fix the program at just that point.

FWIW, in full disclosure, I like the Waves guys, have written manuals for the company and helped set up sites in Nashville when they were generating IRs for their surround stuff. There's nothing like scooting on a huge Genelec across a hockey rink right after the Zamboni has smoothed the ice.

Oh yeah, we should probably clarify that this topic rears its head in two cases: with SRC and when sending the stream to a DAC. It's all about predicting what will happen inside the DAC, but since we don't listen digitally (yet, still waiting for that plug to otic nerve patch), it's important to know what the reconstructed waveform will be.

Also, here's a reference that shows where a lot of these current topics originated: http://www.itu.int/dms_pubrec/itu-r/...3-I!!PDF-E.pdf

Recommendation ITU-R BS.1770-2 (03/2011)
Algorithms to measure audio programme loudness and true-peak audio level

The theoretical max level depends on the reconstruction filter order used in a D/A converter. If the filter order is high enough, you can construct a digital signal that achieves an arbitrarily high overshoot in dBTP. However for real-world musical signals an overshoot of more than +1 dB is highly uncommon.

Correct, but if you have to satisfy a system that rejects you for a .1 True Peak clip, then you're still screwed if you don't have a way to check that. If you're dealing with a spec that uses ITU-R BS.1770-2 (03/2011), it's unfortunate that you have to gear up for that, but I don't think it will be uncommon in the future.

VST > RTAS wrapper?

phil p

Actually, the overshooting phenomena was discovered back in the 40's. I have an old IRE (Institute of Radio Engineers) paper somewhere in my archives, where they go into this in detail. They used to refer to clipping back then as "peak chopping". The broadcast processing crowd "discovered" this in the late 70's and then the digital audio crowd "discovered" this in the early 90's. In reality, NONE of them really discovered anything that wasn't already known.

If you were capable of generating a perfect squarewave in your DAW and set its amplitude to -0.001 dBFS, the signal level would be perfectly legal in your DAW. Further on downstream, when all the harmonics have been removed or filtered from the squarewave, only the fundamental would remain and its amplitude would be 4/pi or +2.098 dBFS. Such is the hazard of heavy digital clipping above zero. That's why I suggest setting a safety limiter threshold a couple of dB below zero.

I still think the TV crowd is off base with their -10 dBFS "true peak" limitations. They should be specifying quasi-peak. Some vendors, like Videotek, actually refer to quasi-peak as "peak" in their literature, hence the possible confusion of terminology.

On a practical note, how would I set up a test to see if Voxengo Elephant really does stop intersample overs? How would I see that they were there (or not there)?

phil p

Check the original vs the processed file in Adobe Audition?

Audition because the metering or display can resolve those peaks? Can any other app do that? Would all audio apps (that show wfms) do this?

phil p

Not sure it "resolves" them, but it does use sinc interpolation to reconstruct the waveform, according to various sources. I think SoundForge has/had this capability as well, but I'm not sure. Because Audition is an editor, not a performance DAW (like Logic) it can afford to have a sinc interpolation algorithm that is optimized closer to an ideal sinc interpolation filter (my assumption here). Realtime filters are by definition approximations because they cannot have infinite impulse response in positive and negative directions, so they use windowing. Both the construction of the algorithm and the implementation in the software and graphical representation come into play here, so the results among DAWs are varied indeed.

I just did a very quick and unscientific test by putting Elephant on an old TV mix of mine. The mix was limited to -9 dB FS with Waves L2. Simply setting Elephant at the same thresholds and 8x oversampling gives an extra peak gain reduction of 0.9 dB so clearly this has some effect.

When I am in a studio with some more tools to measure things properly I will do some more tests. Importantly, how does Elephant react to the unlimited signal and what does the processed signal measure at True-Peak? Just inserting Elephant behind the L2 (as my above simple experiment pretty much does) isn't the proper way to test things of course.

Alistair

A proper way to check if a maximizer limits in dBTP domain is to upsample its output and measure the peak levels. The input signal has to contain a sufficient amount of high frequencies (perhaps even some clipping) for this test to work.

Alternatively, you can use the level meter that can display dBTP levels. Audition can't do it, but there are some plugins. Ozone is one of them (if configured to this mode), I've heard that Inspector XL is another one.

You can also check the table at the bottom of this article: Maximizers. The rightmost column shows whether a maximizer can limit in dBTP domain. L2 does not have this feature.

Thanks for the excellent reference. I'm a bit confused about what +/- means in the "analog detection" column.

You state that:

"Analog detection - means that the limiter can detect and limit peak values in the reconstructed analog waveform (also called inter-sample peaks). This allows creating masters that are more compatible with different D/A converters (+)."

That's clear, but what does the "-/+" indicate? I'm assuming that means that it's "partially compatible" or something like that, but wanted to be sure.

It means that there's some issue with implementation of this feature. Some maximizers implement the "upsample-limit-downsample" approach, which is suitable for ensuring that dBTP levels are brickwalled, but is not the best from the sonic point of view. Double resampling does not have to be present in the main signal chain in order to achieve limiting of dBTP levels.

The Sonnox Limiter has a Reconstruction Meter switch to show you the analogue signal after conversion through a DAC. I always have this on for TV. Of course Mr. Frindle would have thought of this.

I do not believe that the otherwise excellent Massey Limiter has the same functionality.

Everyone should read Alexey Lukin's paper that he linked to. He knows his stuff!

Do you have any idea who Lavry is?

Intersample peaks do exist and do matter, because the sound needs to be converted back to analog before you can hear it!

Try the RTAS plugin. It can write a CSV log which will also show true peak levels.

....Do you have any idea who Lavry is?.....

Dan Lavry?

As far as a standardized methodology or technique for implementing true peak measurement, as applied to deliverable sound files, it's defined in Rec. ITU-R BS.1770-2 19, Appendix 12 to Annex 2: "Considerations for accurate peak metering of digital audio signals".

There are a lot of products out there that probably purport to be doing this, but unless they have tested with known accurate instrumentation, you really don't know for certain what they are doing. Personally, I would tend to trust Tektronix WVR and WFM series of instrumentation, since Tektronix was the equipment vendor that originally implemented true peak in the first place. It is also highly likely that the major networks and organizations are also using this type of intrumentation as the arbiter of waveform legality.

I'd be wary of picking something solely on the basis of its upsampling characteristic, because it could have internal phase shifts or other peak perturbing anomalies that would be exaggerating the actual overshoot, and all you would wind up doing is shooting your self in the foot.

me
textbooks
tutorials by the d/a chip makers
nyquist himself

show me a d/a converter that can go above its maximum voltage that was set for max digital input

no real d/a converter can use a sinc function
the real hardware cannot give you any peaks

besides you can turn the d/a output up or down anyway
just use the handy knob and turn it right/left

and get any values you want back in the analog world

okay
please educate me
been through a lot of d/a reference material
have not run across that item yet

granted a lot has happened with filters since i designed them
(read my name - that says it all)
so what is this magic filter you talk about ?

all the d/a circuits i see in use
limit the maximum output to the maximum voltage
that they then assign to teh maximum binary input

and after that the round knob can raise/lower the output
by turning it to the right/left. you can get any level signal you want from inaudible to pegging the transmitters vu meters at +40

as a mathematician after doing engineering i seem to recall
that the overshoot only occured on a set of measure zero
so is this a real problem or theory ?

a lot of the examples people thow out here violated nyquist criteria. you need to band limit the signal both ways.
nyquist didnt say you could screw with the samples in oddball ways and then get its precursor from the analog domain when you do d/a. besides you dont have sinc functions in real circuits. and we approximate a lot of things by not having an infinite signal length.

so if you use a digital filter to bandlimit the digital values and then do d/a do you really get an overshoot?

if you generate a square wave then it is not bandwidth limited. not nyquist valid to expect to convert that signal.

run it through a digital bandlimiting filter and see if daw still likes it. if not then dont send it downstream.

intersample peaks are a myth
what high school seniors know
urban legend
complete myth

show me a d/a converter that can cause a peak
every vendors chips i have found cant do it
all the recent books/tutorials/references/whitepapers
show that it cant happen

maybe you should digital filter your signal and
bandlimit it before sending it to the d/a
nyquist criteria goes both ways

and *if* there were (there are not) any peaks
after d/a then turn the pretty knob to the left
and lower the signal before you ship it out for use

If you know your sampling theorem, at the ADC, the sampled material is band-passed to prevent aliasing. This is the anti-aliasing filter. Similarly at the DAC it's required to have a low-pass filter to prevent aliasing of Fourier coefficients. This is called the Reconstruction Filter.

If you'd deign to read Dan Lavry's paper, it might help.

I'm talking about the windowed sinc low-pass filter - the most popular type of oversampling filter in D/A converters. It definitely has the ability to increase peak levels of the signal.

This was true for non-oversampling DACs, which in fact quite poorly respect the Nyquist theorem (in part of reconstruction accuracy). But nowadays all music DACs are oversampling, and they use low-pass filters in order to comply with a Nyquist theorem. And this is where inter-sample clipping may occur.

Let me give you an example. Generate the following digital signal at 8 kHz sampling rate:


This is a 2 kHz sine tone digitally peaking at 0 dB FS, but its analog levels go up to +3 dB FS. Now put your DAC to the test: play this tone. My DAC outputs a pure sine, not clipped. It means that my DAC has the capacity for accommodating at least 3 dB of inter-sample peaks.

Now convert (resample) this signal to 44.1 kHz using a decent SRC and let it clip (by saving in a 16-bit format). Here's what you get:


Now listen to it. Clearly, it's different: distorted.

The summary: my DAC can successfully play inter-sample peaks up to 3 dB, but not all DACs are created the same. Some of them may clip on the first waveform. This brings the issue of limiting not only digital levels, but also analog levels (which has been described in many papers by TC Electronic).

It's definitely not a myth, but I'd say that the problem is somewhat over-emphasized. If someone's production is getting rejected because there's an inter-sample peak of +0.9 dB, there may be something wrong with their QA standards or their understanding of dBTP measurements.

In the real world of delivery specs and meeting them, especially when new standards throw new rules and limitations into the mix, the problem is binary:
pass or no pass. If you don't deliver what they say in exactly the way they say it, you fail. And unlike graduate school (yes a few of us went and have some grasp of mathematics), they don't grade on a curve.

I think your real argument is with the Recommendation ITU-R BS.1770-2 (03/2011). Perhaps you should write to them about being high school students who believe urban myths. The rest of us will endeavor to pass the new specs the first time they are given to us, using the most accurate and economical means we can find.

Actually any digital waveform, no matter how square, is a legit bandlimited signal (i.e. is a sampling of a certain bandlimited analog signal).