[castleofargh]

Which is exactly what I said earlier. Someone in the chain must violate Nyquist, or mis-use their gear, to cause this phenomenon.



No. That would violate Nyquist theory. Any high-level sample sub-Nyquist will be reconstructed properly. If it's not, there can only be two reasons, (1) MSB clip (embedded in the PCM), or (2) a DAC that's not meeting it's own specification.



If a mastering engineer pushes program to just below 0dBFS, then a properly specified DAC will reconstruct that master properly. If the mastering engineer purposely clips the program, then a properly specified DAC will clip the master in the same manner. If something other happens, then, yes, the DAC is at fault.

no you don't get it. I suck at explaining things so I'm half to blame here, maybe somebody else will know to give a clear explanation.
a DAC will have, let's say 2V for 0dB, 1V for -6dB and so on. in general it wont be able to go above 2V because there is no bit value above 0dB. if 2 samples are set close to zero dB while mixing, but they happen to be on each side of the sine peak of a high frequency, then the analog signal would require above 2V to properly reconstruct that sine peak while still providing close to 2V at the sample point. boom intersample clipping.
in the digital realm, the sine isn't clipped because no sample was clipped. but because the master didn't care to keep some headroom, we could end up with some of the analog signal clipped at 2V.

does that make sense?

 

[cloggins]

Yes, 2V = 0dBFS. Fine. Can't go above 2V. Yes. If two samples are on each side of a HF sine peak so that a higher peak is missed, then the "sine peak" in question is outside of the Nyquist band, or clipped in the PCM source.

Can't have "missed peaks between samples" in a properly band-limited, unclipped Nyquist space. I'm not arguing that the "measured effect" isn't real, I'm simply saying it may be happening for reasons you're not grasping. Repeat after me -- in a properly band-limited Nyquist space, there are no peaks "between samples" that are missed. What you're talking about is an artifact of poor DAC filtering, most likely.

 

[cloggins]

Nyquist sampling theorem ... has nothing to do with digital quantization,

My friend, I'm sorry. That statement shows an utter lack of understanding. I've been coding DSP for decades. Nyquist (and later Shannon) has everything to do with digital quantization of analog signals.

If an "inter-sample peak" occurs in a digital system, something is broken. Such a signal has not been properly filtered, aliased, "anti-imaged" etc.. Oddly, we've both repeated ourselves now about 3 times, so one of us is not getting it.

 

[bigshot]

Oddly, we've both repeated ourselves now about 3 times, so one of us is not getting it.

Welcome to Sound Science!

 

[ev13wt]

WHile what he writes is true, is usually never very adible. If you set your levels right, its never a problem.

This is why in recording stuff, the extra range from 24bit makes sense. In playback its irrelevant - but more is better so it sells.

 

[gregorio]

If an "inter-sample peak" occurs in a digital system, something is broken.

No, you don't appear to understand what an inter-sample peak is! You actually have it backwards, inter-sample peaks exist precisely because the system is operating correctly in accordance with Nyquist! For example, let's say we have two sample points which correspond to -0.1dBFS but those sample points are either side of the waveform peak. Apply a sinc function during reconstruction and the waveform will be accurately reconstructed, according to Nyquist, with the peak above the analogue equivalent of 0dBFS!! However, the problem is not just limited to analogue headroom in a DAC after reconstruction, it will also occur during a sample rate conversion. If we double or triple the sample rate, say convert from 44.1 to 192, the higher sample rate will now have sample points closer to or on the waveform peak, which fell between our sample points in our example, and those sample points will now have illegal values (>0dBFS). This, unwittingly, is what @bigshot was referring to, AAC and other lossy codecs do not increase the volume but as part of their encoding they do significantly over-sample, thereby realising the peaks between the sample values and clipping them.

The solution has been around quite a few years and has been implemented in the broadcast world, where peak levels are specified in dBTP (dB True Peak) not dBFS (dB Sample Peak) and true peak limiters are employed, not sample peak limiters. The AES has also released specifications for internet distribution and digital radio also in dBTP but there's no legal requirement and the music industry is still pretty much stuck on sample peak limits not true peak.

This is why in recording stuff, the extra range from 24bit makes sense.

16 bit, 24 bit, 500 bits, it makes no difference at all to the issue of inter-sample peaks!

G

 

[gjfs]

Yes, 2V = 0dBFS. Fine. Can't go above 2V. Yes. If two samples are on each side of a HF sine peak so that a higher peak is missed, then the "sine peak" in question is outside of the Nyquist band, or clipped in the PCM source.

It is not out of the Nyquist band. We're talking about amplitude, not frequency.

Nyquist (and later Shannon) has everything to do with digital quantization of analog signals.

No. Go learn the difference between discrete-time systems and digital systems.

Oddly, we've both repeated ourselves now about 3 times, so one of us is not getting it.

That would be you. Go read more about intersample peaks until you understand what they are. Has absolutely nothing to do with "lost information between the samples". You're over-enthusiastically talking about sampling theorem reproducing the input signal exactly, when nobody is saying it does anything otherwise.

 

[Arpiben]

ITU-R BS 1770-4 Annex 2 & Appendix 1
will provide further details about True Peak metering of digital audio signals.
There is a good description of peak sample meters (non true) anomalies.

Anyhow the fs/4 pure tone case used for describing ISP is the worst one.

 

[castleofargh]

ITU-R BS 1770-4 Annex 2 & Appendix 1
will provide further details about True Peak metering of digital audio signals.
There is a good description of peak sample meters (non true) anomalies.

Anyhow the fs/4 pure tone case used for describing ISP is the worst one.

thank you for that. my attempt at making up a physical case with voltage limit, didn't help. but this is as explicit as it gets.

 

[cloggins]

Let's say we have two sample points which correspond to -0.1dBFS but those sample points are either side of the waveform peak.

This is a misconception. If the signal in question is within the Nyquist band, and is not clipped in the original PCM (i.e., two or more consecutive FS MSB samples), there will be no "sample points on either side of a waveform peak." That would be an impossibility IF the path is working properly. Two consecutive PCM FS sample points is defined as a clip. At this point, you have a homework exercise. You must find a valid example of two consecutive FS PCM MSB samples, within a Nyquist band, that is not defined as a clip. I'll wait.

The solution has been around quite a few years and has been implemented in the broadcast world, where peak levels are specified in dBTP (dB True Peak) not dBFS (dB Sample Peak) and true peak limiters are employed, not sample peak limiters.

In the analog domain, a peak can occur at any frequency (obviously). But once a signal has been quantized, it is band-limited to 0.5 the sample freq, and there is no frequency within this range that will be "missed between samples" unless the original PCM source is clipped.

 

[gregorio]

[1] This is the first misconception. If the signal in question is within the Nyquist band, and is not clipped in the original PCM (i.e., two or more consecutive FS MSB samples), there will be no "sample points on either side of a waveform peak." Two consecutive FS sample points is defined as a clip. This is the fourth time I've tried to convey this simple 80-year-old law. At this point, you have a homework exercise. You must find a valid example of two consecutive FS PCM MSB samples, within a Nyquist band, that is not defined as a clip. I'll wait.

[2] In the analog domain, a peak can occur at any frequency (obviously). But once a signal has been quantized, it is band-limited to some upper-limit frequency, and there is no frequency within this range that will be "missed between samples" unless the original PCM source is clipped.

1. You're going to be waiting an awful long time ... until you do your homework first!! And, you've got quite a bit to do, both on the history of Nyquist/Shannon and what it actually means!
The sample points can occur anywhere WITHIN the waveform, not I nor anyone else here is talking about frequencies in excess of the Nyquist point but about parts of the waveform (within the Nyquist limit). The absolute peak point of a waveform (within the Nyquist limit!) my not fall precisely on a sample point, it may fall between the sample points and therefore, while the sample points maybe entirely within the limits, the peak of the waveform may exceed 0dBFS on reconstruction or when up/over sampled. In the example I gave you, the sample points were NOT full scale, they were at -0.1dBFS!

2. The fundamental fact that you appear to be missing/ignorant of, is that the Nyquist point is a limit on the maximum frequency. At 44.1kS/s we can't have any frequencies higher than 22.05kHz, BUT we are not talking about entire waveforms which fall between the sample points (waveforms which would have to exceed the Nyquist limit), we are talking about ONLY the peak part of a waveform falling between the sample points (not the entire waveform!), the peak of entirely legal waveforms which are BELOW the Nyquist limit! Just to be absolutely clear; a sine wave has a cycle which includes a trough and a peak. Somewhere within that cycle we need at least 2 sample points but those two (or more) sample points can be anywhere within that cycle for it to be (legal and) perfectly reconstructed. Those two (or more) sample points do not have to be at the max peak or max trough points (and often aren't) and this is why there is such a thing as intersample peaks and why they exist in entirely legal signals!!

I'm not sure if you don't have any clue about how digital audio works or if you just don't understand what intersample peaks are but either way, before you go shooting off for others to do their homework, you'll look a lot less ignorant/foolish if you do your homework first!! If you don't understand something, then ask, we're happy to try and explain but insulting others because of your own ignorance/misunderstanding won't result in a happy time for you here!

G

 

[cloggins]

it may fall between the sample points and therefore, while the sample points maybe entirely within the limits, the peak of the waveform may exceed 0dBFS on reconstruction or when up/over sampled

Impossible. If there is unclipped PCM data "between sample points" that is "missed" -- then that information is outside of the Nyquist range. That's Nyquist 101.

Now, if there are two or more consecutive PCM FS MSBs, or poorly designed reconstruction filters, or any other part of the processing chain is not perfect, then, yes, there could be clipping problems.

I'm not arguing against the phenomenon of inter-sample clipping, I'm arguing that you're not presenting the causes accurately.

The idea of "missed data between sample points" is a fallacy in a properly working Nyquist system. In a properly functioning Nyquist space, all data is represented faithfully. Not some. Not 99.99%. ALL.

We've covered the possibilities, you can pick whatever makes the most sense.

1.) clipped PCM source (mastering engineer, etc.)
2.) improperly working DAC (bad anti-imaging filter, etc.)

etc...

 

[gregorio]

Impossible. If there is information "between sample points" that is "missed" -- then that information is outside of the Nyquist range.

That answers my previous question, you apparently don't have a clue how digital audio works. That means you've got a great deal of homework to catch up on!

The whole point of the sampling theorem and of digital audio is that you don't need an infinite number of sampling points and therefore, most of the information is between the sampling points! HOWEVER, just because the information is between the sampling points does not mean it's "missed", it can be reconstructed, which is the whole point of the Sampling Theory and why intersample peaks exist!! Again, rather than trying to use Nyquist/Sampling Theory as some sort of insult/weapon, you'd be far better served by actually learning the basics of what it means and how it works!

G

 

[cloggins]

most of the information is between the sampling points!

Incorrect, my friend.

ALL (not "most") of the information is "between the sampling points" and ALL of that information (in theory) can be perfectly reconstructed. There is no "special case" (in valid Nyquist space) where information is "lost" or "indeterminate" between sample points. You need to go back and drill this into your head. Say it over and over: In a properly designed Nyquist space, there is NO information (none, nada, zilch) that cannot be PRECISELY reconstructed in analog space. Go ahead, say it over and over until it sinks in. It's OK, you're among friends.

 

[gregorio]

Incorrect, my friend. ALL of the information is "between the sampling points" and ALL of that information (in theory) can be perfectly reconstructed. There is no "special case" (in valid Nyquist space) where information is "lost" or "indeterminate" between sample points. You need to go back and drill this into your head. Say it over and over: In a properly designed Nyquist space, there is no information that cannot be PRECISELY reconstructed in analog space.

Oh dear, you just don't won't to learn do you? You don't want to learn how digital audio works and you don't want to learn how not to make yourself look foolish/ignorant! If all the information is between the sampling points, then that would include the peaks of the sine waves would it not? Come on, this isn't hard, just go and look up some digital audio theory 101 if you can't get it!

G