Jitter Correlation to Audibility

Quote:

robertsong said:

From Steve Nugent (a leading authority of digital audio). It's a little dated, but there's some fantastic info here. I "bolded" what I found to be the most significant parts.
 
[size=x-small]Jitter Correlation to Audibility[/size]
 
[size=x-small]The correlation of jitter measurements to audibility is in its infancy IME. The problems start with the characterization of jitter. Generally, manufacturers of crystal oscillators specify jitter in terms of RMS jitter amplitude. The problem is that they often neglect to state that this is specified at 10kHz and higher. There is also no spectral or frequency content information specified. This makes it very difficult to tell which oscillators will have audible jitter or objectionable jitter.[/size]
 
I have challenged Steve and several others who maintain the jitter paranoia position on many occasions starting several years back to confirm the claims about the damage of low level jitter by doing better controlled tests none have taken up the challenge the published data stretching back to a 1974 BBC paper never put the threshold of any type/spectra below about 10ns !
 
 
[size=x-small]For instance, Empirical Audio uses two oscillators that are both specified at 2psec RMS jitter. The two oscillators sound radically different to me when used in a re-clocker in a resolving audio system. [/size]
 
 
This is an extraordinary claim yet when Steve participated in public tests using 5 tracks that varied in jitter from 0 to 100ns (0,10,10,30,100) Steve got zero in the correct order 

http://hddaudio.net/viewtopic.php?id=63​

 putting the worst (100ns) track bang in the middle and identifying the 30ns track as having least jitter and categorizing the 10ns track (2nd best ) as the worst, and identifying one 10ns track as being worse than another identical 10ns track. In the same test I managed to correctly identify the order of all 5 (though I guessed the position of two of them that had the same 10ns) of course I cheated by doing a spectral analysis.
 
Even 100ns of jitter did remarkably little actual damage in terms of added distortion. In fact only 1 person in the whole world came forward to prove he was able to reliably correctly identify the 10ns sample and few even managed it with the 100ns sample, I certainly did not , but then I am not claiming uber ears !
 

Click to expand...

Thanks, Nick. You da man.

Quote:

robertsong said:

From Steve Nugent (a leading authority of digital audio). It's a little dated, but there's some fantastic info here. I "bolded" what I found to be the most significant parts.
 
[size=x-small]Jitter Correlation to Audibility[/size]
 
[size=x-small]The correlation of jitter measurements to audibility is in its infancy IME. The problems start with the characterization of jitter. Generally, manufacturers of crystal oscillators specify jitter in terms of RMS jitter amplitude. The problem is that they often neglect to state that this is specified at 10kHz and higher. There is also no spectral or frequency content information specified. This makes it very difficult to tell which oscillators will have audible jitter or objectionable jitter.[/size]
 
[size=x-small]For instance, Empirical Audio uses two oscillators that are both specified at 2psec RMS jitter. The two oscillators sound radically different to me when used in a re-clocker in a resolving audio system. This leads me to believe that the spectrum, or frequency content of the jitter is as important or maybe even more important than the amplitude. I also believe that correlated jitter or jitter with a relationship to the data pattern or audio signal is also more audible than random jitter. This seems to be the consensus in a number of AES papers.[/size]
 
[size=x-small]Studies by the AES (analysis, not human testing) conclude that these are the thresholds of audibility:[/size]
[size=x-small][1] 120psec P-P jitter audibility threshold for 16-bit DAC and 8psec P-P jitter audibility threshold for 20-bit DAC[/size]
[size=x-small][2] 20psec P-P of data-correlated jitter audibility threshold at certain frequencies and "A simple model of jitter error audibility has shown that white jitter noise of up to 180psec P-P can be tolerated in a DAC, but that even lower levels of sinusoidal jitter may be audible"[/size]
 
[size=x-small]Since many measurements (that don't specify any particular frequency content) performed by Stereophile in [3] are above 150psec or close to this, I do not believe that we have reached the limits of jitter audibility yet. I suspect that P-P jitter needs to be almost an order of magnitude smaller, or around 15psec to be inaudible in all systems.[/size]
 
[size=x-small]I believe the ability of the human ear/brain, particularly the trained ear, to hear minute differences, particularly data-correlated jitter, is grossly underestimated. The live listening AB/X studies published to date (that I have read) are inconclusive IMO. The systems used were not resolving enough IMO, the recording quality was not good enough and the test signals were random and not correlated and therefore inadequate to properly test for jitter audibility. I tend to believe the numbers arrived at by the AES analytical studies rather than the A/BX listening tests. [/size]
 
[size=x-small]There are a series of double-blind tests being performed by many audiophiles using synthetic jitter tracks provided by HDTracks. These may shed some new light on true audibility. Again, the effectiveness of these experiments is only as good as the quality of the tracks provided, the jitter that was synthesized and the audio systems that are used for testing. The results from the first set of jitter tracks shows just how unresolving most audiophile systems are. There are couple that could pick out the majority of the tracks by increasing jitter, but the majority could not hear any difference between the tracks, even though the jitter ranged from 0 ns to 1000ns I believe.[/size]
 
[size=x-small]Another interesting thing about audibility of jitter is it's ability to mask other sibilance in a system. Sometimes, when the jitter is reduced in a system, other component sibilance is now obvious and even more objectionable than the original jitter was. Removing the jitter is the right thing to do however, and then replace the objectionable component. The end result will be much more enjoyable.[/size]
 
[size=x-small]Jitter can even be euphonic in nature if it has the right frequency content. Some audiophiles like the effect of even-order harmonics in tubes, and like tubes, jitter distortion can in some systems "smooth" vocals. Again, the right thing to do is reduce the jitter and replace the objectionable components. It is fairly easy to become convinced that reducing jitter is not necessarily a positive step, however this is definitely going down the garden path and will ultimately limit your achievement of audio nirvana.[/size]
 
[size=x-small]Sibilance in a system caused by preamp, amps and other components and cables can also be so high that changes in jitter are not very audible. This is why there is such contention on the web forums about jitter and its importance. What matters in the end is if you are happy with the sound of your system, and whether or not you can hear this distortion.[/size]

Click to expand...

Let's take a look at Nugent's interpretation of the "consensus" selectively drawn from two papers. Unfortunately I can't breach the paywall of the first cite, so let's focus on the extraordinary claim of 20ps jitter being audible that Nugent draws from the second paper, available at http://www.scalatech.co.uk/papers/aes93.pdf.
 
Here's the linchpin to his argument:
 

 
Wow! That simulated jitter touches the other curve! Case closed!
 
Wait a second...
 

 
Wow! That dither also touches the other curve! Everyone can hear 16 bit dither!
 
Hold on, earlier in that sections it states...
 

 
But wait, didn't they just not do that? Their simulated jitter is at a lower level than their dithered noise!
 
There's more:
 

 
Prior to the introduction of the graphs the authors warn us of two assumptions:
 

• Their simulation of jitter must intersect the absolute threshold of human hearing to be deemed audible.
• Their simple model of audibility "will yield pessimistic results" because it in no way accounts for masking.

 
Now these two are some pretty major assumptions. To begin with, the minimally audible field is measured by putting subjects in an anechoic chamber. As in: people with good hearing in a completely unnatural environment will react to these minute noises. This has nothing to do with listening to music in any sort of real life scenario. Further, masking is a pretty big deal, it makes all the difference in the world when it comes to listening tests. It might be OK to exclude it in a simple model used to prove a point in a technical paper, however it is not OK to claim that this model will pan out in the real world with a signal playing.
 
Nugent ignores these limitations, he gleefully titles this section "Jitter Correlation to Audibility" despite the citations being models and not extensive controlled tests subject to statistical analysis. He proceeds to claim that not limiting these levels of jitter will prohibit one from entering "audio nirvana" (it would be funny if this were not a direct quote).  We are also treated to the irony of him claiming that "[t]he live listening AB/X studies published to date (that I have read) are inconclusive IMO"; whereas cherry-picking a calculation derived from a simplified model is presumably conclusive enough for his article.

Quote:

mikeaj said:

Those figures are well known (and often cited by some), but keep in mind they're theoretical and based on conservative calculations for what should be masked and inaudible based on other research on masking effects.
 
Sometimes listening tests aren't as sensitive as they could be, sure. There could be more going on, but there could always be more going on. If you're going to throw out the listening tests run on the subject, then where are you going to start? Either the basis of concern is some conjectures or it's anecdotes. I don't dismiss it at all as an impossibility, but all the caveats and concerns quoted above have the same structure as the usual FUD arguments.

Click to expand...

 
Thanks Mike for your input. I totally see where are you coming from. But why is that you and I see this as inconclusive, while others do not? How is it that some people (in this thread even) come to the conclusion that jitter is inaudible?

For me, thresholds of human perception in real world music listening tests established by the AES are enough. Other people love to think about all the "what ifs" that end up being a bunch of discussion about angels dancing on the heads of pins. This is the problem with the discussion of jitter. People spend weeks discussing trivia and a couple of seconds discussing real world facts. It confuses people to spend a lot of time dotting is and crossing ts. I can understand how you could be confused.
 
You have gotten some mighty good evidence in this thread that jitter isn't an issue, and for you they are still "inconclusive". What would it take to convince you?

Quote:

xnor said:

So we have theoretical tests with special test signals and listening tests with music. The latter show that jitter isn't audible unless you go over several thousand ps.
 
The "high" jitter in HDMI receivers also was fixed quite some time ago and some can achieve incredibly low values of just a few tens of ps.
 
 
=> No point in spending more money to get less jitter.

Click to expand...

 
 
You sound quite certain of this.

Quote:

bigshot said:

 
You have gotten some mighty good evidence in this thread

Click to expand...

 
 
Where?
 
 
Quote:

What would it take to convince you?

Click to expand...

 
I'm starting to see more and more that it can't be proven either way. Steve N cited some examples where the test could be flawed. Also this test from Antepode- http://www.scalatech.co.uk/papers/aes93.pdf

Also: USB>SPDIF converters. If it isn't jitter that makes them sound so different from model to model and compared to built in usb receivers in DACS, then what is it??
 
Ditto for SPDIF cables.

Did you look at the papers I posted? Did you compare the thresholds of audibility they establish to the levels of jitter in typical audio components? Have you made any attempt to wrap your head around how much a picosecond and a nanosecond is? Do you know what -80dB sounds like in the real world? Do you know how masking works?

You're going at the problem from the wrong end. You are assuming jitter is a problem and you're looking for evidence of that. Instead of doing that you should just try to figure out what jitter is, how big it is, and how that would affect listening to music. Don't assume that because you think there is a difference, it must be jitter. Break it down and analyze each possibility. Look for what is likely and work past what isn't.

I completely understand how someone could glaze over at all this stuff and just assume what they feel in their gut is real. I completely sympathize. I was there once too and it was a chore to figure out what all this stuff really means. Yes, it's needlessly dense. But work through it. Do the homework. Google. Figure out the real truth, not just what you want to be the truth.

.

Quote:

robertsong said:

Thanks Mike for your input. I totally see where are you coming from. But why is that you and I see this as inconclusive, while others do not? How is it that some people (in this thread even) come to the conclusion that jitter is inaudible?

Click to expand...

 
I mean this in the sense that there's almost nothing that's absolutely conclusive. Also, based on one study showing 10 ns requirement for sinusoids, that is not a lot of evidence by itself (though it seems to be backed by some other data, listening tests that I am not as familiar with). Maybe the threshold is 5 ns under some circumstances? etc. Or 10 ns may be too low already. You can't be very confident on an exact figure without more studies. One study or two studies aren't irrevocable truth. It's always possible that some studies aren't done properly, or some conditions X weren't met, or whatever. Even if it's peer-reviewed by AES, it could be wrong. Like I said before, though, these are the usual weasel-out FUD arguments, and it's not like people should just assume that the studies are bunk...
 
I don't think there's an argument here that very high levels of jitter is inaudible. There are synthesized examples showing audible levels (not necessarily the lowest possible levels necessary to be audible). These are much higher than found in most devices, even many cheap audio devices.
 
As for Nugent's line of analysis, that is, as I stated earlier, "conjecture". Like I said, if you throw out the listening test data, then you can't make an argument for low levels of jitter being audible without relying on some anecdotes or the guesswork analysis. Actually, I was not aware of exactly where the figures came from, so even before when I labeled it as such, maybe I was giving it too much credit. I thought that the analysis assumed masking, whereas it appears that it doesn't. So my original opinion was based on faulty assumptions.
 
 
So the current position is that there's no credible listening tests that support any claim for jitter audibility at levels achieved by most cheap audio devices. There aren't extensive tests for a lot of things. You're not going to find reams and reams of directly-applicable research on many things, especially in cases where there's not much anything interesting to report. And theoretical arguments for lower levels of audibility are based on stringing together the most conservative figures together to reach the lowest possible number. As in many places in audio, this is the place where some people are just going to be skeptical unless you can show them new data. As for the others, I don't think there is solid ground to stand on (...yet, or ever, depending), unless you consider conjecture and anecdotes to be bedrock.
 
Was the purpose of the original post to ask a question or to convince people? Maybe I'm misinterpreting the motives and this situation.
 
As for hearing differences, that'd be going back to anecdotes. Which doesn't really get us anywhere. You know how easy it is to be presented with identical stimulus and perceive different things, right? If presented with almost identical (functionally identical, and even if impossible to distinguish) stimulus, you can also perceive different things.

Quote:

bigshot said:

Did you look at the papers I posted? Did you compare the thresholds of audibility they establish to the levels of jitter in typical audio components? Have you made any attempt to wrap your head around how much a picosecond and a nanosecond is? Do you know what -80dB sounds like in the real world? Do you know how masking works?

Click to expand...

I think it might be worth spelling it out for those playing along at home:
 
I can easily hear a -96dB signal in my gear: I tell Windoze to send 16bit audio to my Modi, crank my Magni to 11, and hear the 1bit noise from Kmixer's dithering which lies at -96dB full scale. Of course, I can't have any actual music playing, because it would blow my eardrums through my skull if there was anything remotely close to 0dB full scale playing at the same time.
 
Jitter can only have an effect when signal is present. A jittery sequence of 0's still perfectly reconstructs the zero-amplitude signal. For that jitter to be present at -80dB full scale, the actual signal level has to be at 0 dB full scale. If the signal is below clipping levels, say -10 dB full scale, then those jitter side lobes are at -90 dB full scale.
 
If you guys want to hypothesize worst case scenarios when jitter could be most audible, I would be happy to do the sample calculation or create an artificial waveform with and without jitter of a specified size and spectral content.
 
Cheers!

Quote:

robertsong said:

Also: USB>SPDIF converters. If it isn't jitter that makes them sound so different from model to model and compared to built in usb receivers in DACS, then what is it??
 
Ditto for SPDIF cables.

Click to expand...

 
That's a good question on the converters.
 
An audio-nihilist says it's all 1's and 0's, the power light is either on or off.
 
I'm not sure how $500+ plastic boxes would sell if they all really do sound identical.
 
The best USB chip available and extremely low measured jitter results in a USB>SPDIF converter shouldn't cost more than $100 though I think, the rest is retailer mark-up etc.

Quote:

robertsong said:

Also: USB>SPDIF converters. If it isn't jitter that makes them sound so different from model to model and compared to built in usb receivers in DACS, then what is it??
 
Ditto for SPDIF cables.

Click to expand...

 
If they were not compared under controlled conditions (which I suspect is the case), then this is fairly easy to answer.

Can you hear this simulated jitter (click on the graph to zoom) ? The +/- 3 kHz sidebands alone are more than 10 ns (p-p) jitter, so it should be easy to hear for any golden ears. Post a 30-second part of the music you want to test, and I will create a version with jitter.
 

It is of course also possible to modify the parameters of the jitter, or make the model more complex (more sidebands, different filtering of the noise components, etc.). It currently includes:
- mono frequency modulation by 3 kHz tone + lowpass filtered (brown) noise
- stereo phase modulation by filtered noise

Quote:

Originally Posted by mikeaj /img/forum/go_quote.gif
 
How is the timing usually reported for jitter, anyway? Max difference of one cycle and another? Average difference between average time and min, between average and max? Some kind of rms value? Anyway, with a 10 kHz signal, a cycle takes 100,000 ns. Even if 10 ns were to indicate that some cycles take 100,010 ns and others take 999,990 ns, you'd expect to see spikes or spreading at 9990 - 10010 Hz, right? What are those spikes doing at 7 kHz and 13 kHz then? That's a bunch of orders of magnitude away.
 
Maybe I should think about this after going to bed.

Click to expand...

 
The ns and ps values for jitter are peak to peak levels for phase (delay) modulation. For modulation by a sine, the formula is
 
Jitter time (peak to peak) = sideband level / (tone frequency * ℼ / 2)
 
For example, 7.66 ns p-p jitter by a sinusoid creates a pair of sidebands at 0.00000000766 * 11025 * 1.5707963268 = 0.00013266 = -77.5 dB level relative to a 11025 Hz fundamental.
 
Edit: sometimes it is specified as a peak (rather than peak to peak) level, which halves the number. That is the case with the above paper that claims the theoretical audibility of 20 ps peak (= 40 ps peak to peak) jitter under worst case conditions.
 
Quote:

Originally Posted by mikeaj /img/forum/go_quote.gif
 
By the way, those are spectrum from J-test: 11025 Hz single tone (quarter of 44.1 kHz, tested at that sampling rate) with LSB toggled. It's supposed to be a worst-case scenario.

Click to expand...

 
The LSB toggling is for interfaces like S/PDIF where the binary audio data can potentially modulate the clock. So, a 16-bit JTest signal looks like this in hexadecimal:
  0x4000, 0x4000, 0xC000, 0xC000,
  ... (repeat the above for a total of 96 samples)
  0x3FFF, 0x3FFF, 0xBFFF, 0xBFFF,
  ... (repeat the above for a total of 96 samples)
In other words, there is a 44100 / 192 = 229.6875 Hz square wave at a peak to peak amplitude of 1 LSB added at 44.1 kHz sample rate. In a badly implemented S/PDIF interface, the audio data bits modulate the clock to some extent, which will create sidebands in the analog output signal.