Audibility of Jitter, Proposing a series of tests Options

[cabbagerat]

The threshold of audibility of phase noise in ADC and DAC clocks is a fairly contentious issue in the HiFi and audiophile world. Some sources claim that jitter is clearly audible at low levels, and some claim that high levels of jitter are inaudible. The literature describes several tests, many with conflicting results.

One of the chief difficulties in testing the audibility of jitter is that it requires a complex hardware setup, which means that many listeners would be required to be present for an time consuming (and expensive) on site test. Over the last couple of months I have been thinking about organising a distributed listening test to look at the audibility of jitter in audio applications, based on algorithms for simulating the effects of jitter on signals. These algorithms are fairly well described in RF and telecomms engineering literature, and would be interesting for comparison purposes.

The kind of thing I have in mind is this:
Use samples which are accepted to sound good -> simulate jitter -> perform listening tests -> perform more tests at different levels of jitter depending on results

The purpose of this thread is to get ideas of the Hydrogenaudio community about performing these tests. Some of the things I would appreciate input on are:

• Would such a test be useful
• Would the results of such a test be valid?
• What sort of test procedure would be best?

[Gigapod]

The threshold of audibility of phase noise in ADC and DAC clocks
...

I would suggest you begin with a working definition of "Jitter", which you could post to the HA Knowledgebase. Is jitter "phase noise" ? What does it sound like? What does it do to a sine wave that goes through ADCs and DACs ? Does it matter at all ? What is the order of magnitude of jitter in PCs? and in high-end audio equipment?

This post has been edited by Gigapod: Dec 27 2006, 16:37

[dariju]

There is a good article about jitter @ Digital Domain:

http://www.digido.com/modules.php?name=New...icle&sid=15

I'm looking forward to see such test... How many of you believe in "jitter effect"?

[cabbagerat]

I would suggest you begin with a working definition of "Jitter", which you could post to the HA Knowledgebase. Is jitter "phase noise" ?

A good point. I would go with a definition of Jitter as the high frequency component of the time interval error in the clock signal. Where the cutoff of "high frequency" is must be chosen, as must the spectrum of the modelled noise. This spectrum is one very hard problem - it's likely to consist of a variety of intermodulation products, PSU noise, other noise, harmonic distortion and vary widely between different pieces of equipment.

What does it sound like? What does it do to a sine wave that goes through ADCs and DACs ? Does it matter at all ?

Jitter spreads the spectrum of a sine wave out, effectively by convolution with the spectrum of the jitter. I have no idea what it sounds like, except in some extreme cases. And whether it matters at all is one of the questions I am hoping this test will help answer.

What is the order of magnitude of jitter in PCs? and in high-end audio equipment?

I have no idea about PCs, but total jitter figures for midrange HiFi stuff is generally somewhere between 0.1ns and 100ns - a three order of magnitude range.

[Gigapod]

There is a good article about jitter @ Digital Domain:

http://www.digido.com/modules.php?name=New...icle&sid=15

I'm looking forward to see such test... How many of you believe in "jitter effect"?

Having just read the article you linked to, I found it a little too much "audiophile"-minded and not very clear about the technical issue of jitter. Quoting: " ...The sonic results of passing this signal through processors that truncate the signal at -110, -105, or -96 dB are: increased "grain" in the image, instruments losing their sharp edges and focus; reduced soundstage width; apparent loss of level causing the listener to want to turn up the monitor level, even though high level signals are reproduced at unity gain..."

I know quartz oscillators have a small amount of phase noise (jitter).
How much does the phase noise in a 12MHz quartz oscillator driving a DAC at 48kHz affects the final waveform?

[Gigapod]

... total jitter figures for midrange HiFi stuff is generally somewhere between 0.1ns and 100ns ...

100ns sound a little bit high (for a 10MHz quartz oscillator that's one full cycle) but OK, let's work with that.
How would a 10MHz signal modulation of an audio signal ranging from 20Hz to 20kHz affect the signal?
It seems all the intermodulation products would be way above the audibility range, no? Or am I missing something?
I think a typical figure in an PC-grade audio codec would be 1 ns jitter (approx. 1% jitter in a 10MHz clock), which would represent approximately 0,005% jitter in a 44.1kHz clock (if my napkin calculations are right).
My a priori guess is the ABX test would be unconclusive...

(edit) I just checked the ALC882 codec datasheet and the 24MHz bit clock input is specified with a maximum jitter of 2ns.
The S/PDIF-OUT jitter is specified at 4ns max (6.144MHz clock). I think the S/PDIF-OUT clock is PLL generated internally in the codec.

This post has been edited by Gigapod: Dec 27 2006, 18:59

[Pio2001]

• Would such a test be useful
• Would the results of such a test be valid?
• What sort of test procedure would be best?

Yes it would be useful. Jitter is the main cause invoked in order to explain the alledged sonic difference between the digital output of a 50 € DVD player and a 10,000 € dedicated digital drive.

The results would be valid if several conditions are fullfilled.
On top of my head I can think about
-The kind of jitter. Some kind might be audible, some kind inaudible. We must choose the kind of jitter that should cause the biggest audible effect.
-The test must be performed on a system whose jitter is small compared to the tested amount of jitter.
-The samples used must be sensitive to jitter.

A full scale 20 kHz sine with jitter introducing a 3.5 kHz artifact should be the most sensitive combination.

The most difficult part in this test, in my opinion, will be to get comprehensive jitter analysis from consumer CD players. Especially if jitter is signal dependant.
Jitter should also be analyzed directly at the clock output in order to account for hardware induced jitter. Some claim that heavy error correction or tracking corrections burdens the power supply, which could in turn affect the clock stability.

This post has been edited by Pio2001: Dec 27 2006, 18:39

[Gigapod]

...
Jitter should also be analyzed directly at the clock output in order to account for hardware induced jitter.
...

I think you meant "measured" in the phrase above. However, measuring jitter is not a trivial thing, whether we are measuring 0.1ns or even 10ns.
I would dare say that none of us HA readers has the equipment to directly measure jitter with any degree of precision.

[cabbagerat]

The results would be valid if several conditions are fullfilled.
On top of my head I can think about
-The kind of jitter. Some kind might be audible, some kind inaudible. We must choose the kind of jitter that should cause the biggest audible effect.
-The test must be performed on a system whose jitter is small compared to the tested amount of jitter.
-The samples used must be sensitive to jitter.

All three of these are difficult problems - and would need careful test design to handle. Possibly starting with something along the lines of "is 1us of jitter audible? in test samples? in music?" and moving down to the harder problems if that test has a positive result would be a good plan. I don't know what the right solution is.

I would dare say that none of us HA readers has the equipment to directly measure jitter with any degree of precision.

I do have access to such equipment when I am at University, and hopefully I can get another student to teach me to use it. Real measurements would be invaluable but, as Pio2001 says, if the jitter is data dependent (and it seems to be in many cases) it becomes a much harder problem.

Gigapod - you sound like you know quite a lot about jitter/phase noise - thanks for your input so far. Keep it coming

[Gigapod]

...
I do have access to such equipment when I am at University, and hopefully I can get another student to teach me to use it. Real measurements would be invaluable but, as Pio2001 says, if the jitter is data dependent (and it seems to be in many cases) it becomes a much harder problem.

Gigapod - you sound like you know quite a lot about jitter/phase noise - thanks for your input so far. Keep it coming

Cabbagerat, I think it's very interesting if you can measure jitter in audio equipment in a well equipped university lab, and later determine through listening tests if it has any audible effect. I think it's always satisfying to debunk a myth with some solid experimental data, and I am highly suspicious of the whole jitter audiophile shebang.
Actually a long time ago I studied precision clocks and did some superficial documentation gathering on quartz oscillators. I think nowadays even the cheapest CD player (like the ones costing 15 euros that you can buy at the supermarket) uses a quartz clock base, because quartz is so cheap. Typically I think the jitter figures will be very low, because even with 25% jitter (highly unlikely) at 10MHz that's only 25ns, but I have no experimental data to offer to support that assertion (lack of equipment, lack of time and vague motivation).
If I may offer some guidance here, I would say, take a single piece of audio equipment (e.g. a normal CD player) to your university, find the quartz oscillator, and check the jitter at the buffered output of the quartz oscillator (not directly at the quartz leads, obviously). Of course schematics would help a lot, but if you can't find the schematics it shouldn't be too difficult to find the quartz oscillator circuitry.
BTW at 10MHz a few pF will significantly affect the waveform, so I recommend a FET active probe with 1GHz bandwidth.
If you find out the jitter is of the order of 1ns, I wouldn't bother with the listening tests...

This post has been edited by Gigapod: Dec 27 2006, 23:48

[Zster]

There is a good article about jitter @ Digital Domain:

http://www.digido.com/modules.php?name=New...icle&sid=15

I'm looking forward to see such test... How many of you believe in "jitter effect"?

Having just read the article you linked to, I found it a little too much "audiophile"-minded and not very clear about the technical issue of jitter. Quoting: " ...The sonic results of passing this signal through processors that truncate the signal at -110, -105, or -96 dB are: increased "grain" in the image, instruments losing their sharp edges and focus; reduced soundstage width; apparent loss of level causing the listener to want to turn up the monitor level, even though high level signals are reproduced at unity gain..."

I know quartz oscillators have a small amount of phase noise (jitter).
How much does the phase noise in a 12MHz quartz oscillator driving a DAC at 48kHz affects the final waveform?

I agree with your assessment that the article is too audiophile. Here is a more scientific approach by someone who has designed ADC and DAC for MRI and military devices and makes studio quality ADC and DAC’s.

http://www.lavryengineering.com/white_papers/jitter.pdf

unfortunately jitter is very real.

[Gigapod]

...
Here is a more scientific approach by someone who has designed ADC and DAC for MRI and military devices and makes studio quality ADC and DAC’s.

http://www.lavryengineering.com/white_papers/jitter.pdf

unfortunately jitter is very real.

Thank you, that was an interesting paper.

I agree with you that jitter is real. However, the author of the paper, Dan Lavry, used FM modulation of sampling rates to simulate jitter. This is definitely not an accurate simulation. Jitter is also called "phase noise" for a good reason: it has a noise profile.

Just a note: your inexpensive PC codec uses a 24MHz clock to drive an internal PLL that itself drives the sampling rate of the A/D and D/A converters. It's next to impossible to FM modulate that internal PLL...

The second graph on page 6 is a more accurate representation of real-life jitter effects, and you can see they are below the noise floor of the signal.

Cabbagerat suggested in a post in another thread that most manufacturers don't quote jitter figures for their audio equipment gear because the effects of jitter can ultimately be measured in noise and THD figures. I think this is correct; I would only add that providing jitter figures directly would be a) difficult (because jitter is difficult to measure and where exactly do you measure it ?) and b) meaningless, as you can only hear the side effects of jitter as added noise or distortion.

Nowadays noise and distortion figures hover well below audible levels, so I confess I am not too worried about jitter, even though it's real... in a sense.

[cabbagerat]

Cabbagerat suggested in a post in another thread that most manufacturers don't quote jitter figures for their audio equipment gear because the effects of jitter can ultimately be measured in noise and THD figures. I think this is correct; I would only add that providing jitter figures directly would be a) difficult (because jitter is difficult to measure and where exactly do you measure it ?) and b) meaningless, as you can only hear the side effects of jitter as added noise or distortion.

The problem here is that THD does not measure how bad the distortion sounds. In general, higher THD is worse, but a tube amp fan will tell you that 2% THD doesn't always sound bad. That's well accepted, but it means that audiophiles (and others) will claim that the distortion introduced by jitter is particularly bad.

Lavry's paper is fascinating. However, I agree with Gigapod that his treatment using narrowband noise and FM modulation is too simplistic to be an accurate representation of what happens in reality - the phase noise on quartz oscillators does not resemble the test signals he used. It's the best audio-specific treatment I have read.

[Gigapod]

...The problem here is that THD does not measure how bad the distortion sounds. In general, higher THD is worse, but a tube amp fan will tell you that 2% THD doesn't always sound bad. That's well accepted, but it means that audiophiles (and others) will claim that the distortion introduced by jitter is particularly bad.
...

Imho when THD is below 0.01% it matters little whether it has only odd harmonics or only even harmonics: you just can't hear it anyways ("audiophiles" will certainly disagree...). ABX testing of power amps (Transistor x Valve) seems to support this.
Similarly, if the added noise is below the noise floor of a recording. However, note that noise spectrum analysis is required: changing the noise spectrum profile can be audible in some circumstances.

Getting back to the paper by Lavry:
FM modulation of the sampling rate -> I.M. distortion byproducts.
Phase noise (jitter) in the sampling rate -> noise byproducts.
Nothing new, really.

Unfortunately Lavry provides very little in the way of experimental data and perhaps I missed it but I didn't find any mathematical analysis. I am sure one can easily deduce a mathematical formula that will relate the amplitude and frequency of a signal, jitter (or FM modulation) in the sampling rate and noise/IM distortion levels in dB (or %).

I wouldn't be surprised if normal jitter found in commodity-priced audio equipment would have a noise effect at the quantization noise level in 16-bit CD audio. 24-bit audio probably requires a lot more care with jitter. But then 24-bit audio requires a lot more care with everything, doesn't it?

[wimms]

Jitter matters only during slope of the signal. In severe case jitter is causing LSB error. By taking fastest slope, and limiting jitter induced error at 1 LSB, we can find out max acceptable jitter in terms of technical perfection.
For 16-bit system, to keep sample error below 1 LSB, 15KHz signal of 0db can tolerate no more than 324ps of jitter peak.

In other words, we can change the problem into asking how many LSBs in error can we get away with without being audible. Obviously, it does matter whether these errors are white noise-like, signal correlated or alien.

I suggest few papers from:
http://www.essex.ac.uk/ESE/research/audio_...blications.html

C41 IS THE AES/EBU/SPDIF DIGITAL AUDIO INTERFACE FLAWED?, Dunn, C. and Hawksford, M.O.J., 93rd AES Convention, San Francisco, preprint 3360, October 1992

C134 JITTER SIMULATION IN HIGH RESOLUTION DIGITAL AUDIO, Hawksford, M.O.J, 121st AES Convention, San Francisco, October 2006, paper 6864

[Gigapod]

...
C134 JITTER SIMULATION IN HIGH RESOLUTION DIGITAL AUDIO, Hawksford, M.O.J, 121st AES Convention, San Francisco, October 2006, paper 6864

Whoaaa! Some serious math here. That'll take some time to digest, that one!!!!

Thanks wimms!
Should I send you the coffee bill or are you willing to lead us through this paper?

(edit) Your sig -> Very funny!!!

This post has been edited by Gigapod: Dec 29 2006, 16:37

[cabbagerat]

The second one of those paper is fascinating. It's the mathematical treatment of jitter that I have been looking for for a while. One excellent quote from the paper is:

Uncorrelated jitter, although it can result in modulation noise, is generally believed to be more benign compared to jitter that has a correlation with the audio data or has a relationship to a periodic signal such as mains hum.

Much of the complex maths in the paper is derivation of a simplification of the standard bandlimited interpolation formula, with added jitter factors. This optimisation might be necessary for real-time processing, but for offline processing simple truncation can make performance of the unsimplified algorithm acceptable on modern processors. Sections 7 and 8 (one analyzing LPCM and one sigma-delta) are interesting, as are his conclusions.

[krabapple]

The second one of those paper is fascinating. It's the mathematical treatment of jitter that I have been looking for for a while. One excellent quote from the paper is:

Uncorrelated jitter, although it can result in modulation noise, is generally believed to be more benign compared to jitter that has a correlation with the audio data or has a relationship to a periodic signal such as mains hum.

Much of the complex maths in the paper is derivation of a simplification of the standard bandlimited interpolation formula, with added jitter factors. This optimisation might be necessary for real-time processing, but for offline processing simple truncation can make performance of the unsimplified algorithm acceptable on modern processors. Sections 7 and 8 (one analyzing LPCM and one sigma-delta) are interesting, as are his conclusions.

For all its complex maths, there's an unfortunate lack of audibility data in the paper, or even a summary of previous such data, to indicate how 'important' this ends up being to the consumer .... so what does it do towards addressing the subject of this thread?

This post has been edited by krabapple: Jan 13 2007, 00:13

[Gigapod]

...

For all its complex maths, there's an unfortunate lack of audibility data in the paper, or even a summary of previous such data, to indicate how 'important' this ends up being to the consumer .... so what does it do towards addressing the subject of this thread?

Because - unlike for example harmonic distortion - you can't hear clock jitter directly, as explained in the various papers, you can only hear its effects on the data as it gets converted back from digital to analog.

And as I wrote in my rant above (sorry I am repeating myself, must be getting old), the effects (basically distortion & noise) are usually much below others (for example, noise and HD in the microphone that is used to take the recording in the first place, or quantization noise in 16-bit recordings, etc), even in commodity-grade audio equipment.

[krabapple]

well, let me offer this 2005 paper for consideration, then


http://www.jstage.jst.go.jp/article/ast/26/1/50/_pdf


Audibility threshold for timing jitter, for 'golden eared' listeners in a two-alternative forced-choice paradigm using their preferred listening environment and samples: 250 ns.

This post has been edited by krabapple: Jan 13 2007, 04:36

[Woodinville]

well, let me offer this 2005 paper for consideration, then


http://www.jstage.jst.go.jp/article/ast/26/1/50/_pdf


Audibility threshold for timing jitter, for 'golden eared' listeners in a two-alternative forced-choice paradigm using their preferred listening environment and samples: 250 ns.

What was the spectrum of the jitter? That's really a key question, you know.

[Kees de Visser]

For the Dutch speaking and jitter interested readers it might be interesting to know that the Dutch AES has planned an evening about "digital clocks" (wordclock distribution in the digital studio). Non-members are welcome too in Utrecht on 25/01/2007. More info can be found here.
Jitter is one of the topics that will be discussed and demonstrated.

[DualIP]

For tests, why not generate a simple program that can simulate jitter?

The proposed program simulates an analog source represented by in.wav, which is AD converted using a non jitterfree convertor

How this can be done (I think):

command format:
simulatejiitter.exe input.wav jitter.wav output.wav

jitter.wav contains values, that determine how much time-offset should be created for the corresponding input.wav sample. Jitter.wav can be given any shape/spectrum.

For instance. jitter sample=0 equals no time off-set, and jitter sample = 32767 results in max positive time offset (for instance 1 complete sample period)

As an example, output samples can be calculated using formula:
out[n]:=in[n] + (in[n+1])-in[n-1])*jitter[n]/65536


For simplicity, this example uses a simple lineair interpolation for approximating signal value near sample
in[n]. A 2nd order function would be already be more accurate.

Best solution would be infinite FIR/ nyquist lowpass, that can calculate any, in between sample value, but that's not what I call simple!

If out.wav is 24bit format, rounding errors are kept out.

[knutinh]

Correlated jitter is probably worse than random jitter

High-frequency jitter is probably worse than low-frequency jitter

Any simulation of jitter for detectability threshold will be limited by the playback jitter performance - typically unknown.



By doing detectability measurements, coupling those to models from good measurements, one should be able to decide this once and for all.

Any progress in finding jitter attenuation in "typical" DAC/surround receivers?

What about inserting a RLC network along an spdif path to provoke real spdif-jitter performance of various spdif receivers?

-k

[cabbagerat]

For tests, why not generate a simple program that can simulate jitter?

I have a MATLAB program which does this, based on bandlimited interpolation. I would be happy to post the code when I get my home PC out of storage in a few days.