Jitter: It’s a good thing.

[Ulas]

Every audio recording process results in some added noise: Tape hiss, vinyl surface noise, and for digital, quantization noise. Quantization noise, sometimes call sampling noise, comes from having to map an infinite number of analog amplitudes into a limited number of discreet digital values at each sample point. In theory, that noise amounts to +/- ½ LSB; in practice it is probably much larger.

Quantization noise wouldn’t be noticeable is it were truly random because human hearing has a knack for tuning out and ignoring random noise. Unfortunately, quantization noise is correlated with the signal and is hard to ignore. Since it can’t be ignored or eliminated, the next best thing is to try to mask it and that’s the purpose of dither. Dither is random noise that is added to the audio signal to randomize and decorrelate the quantization noise. The trouble is, the dither signal most commonly used is not truly random. It is the output of a computer algorithm that is best described as pseudo-random. Although short strings of pseudo-random numbers appear to be random, the overall sequence is entirely predicable and the sequence will repeat itself over and over again. What’s worse, repeating patterns within the sequence are easily detectable.

Human hearing is well adapted to recognize intelligent audio signals buried in random noise. Two such intelligent signals are speech and music. The hallmarks of both are predictability and repeating patterns. Unfortunately, they are also hallmarks of pseudo-random numbers. In the early days of digital music syntheses, the sounds of percussion instruments were modeled by sampling pseudo-random numbers. You could change the tone color of the sound by changing the constants in the random number algorithm. Similarly, every dither generator has a particular tone color and it’s pretty easy to recognize: Just hookup a DAC to different dither sources and listen.

Just as dither is added to the signal to mask and decorrelate quantization noise, jitter helps mask and randomize the colored noise of dither. Added jitter is analogous to the use of Gaussian blur in digital photography.

While jitter is beneficial for non-oversampled DACs, as described, it is even more vital for oversampling DACs. The interpolation filter attempts to create a sinusoidal wave that incorporates and expands the given data points. Unfortunately, the filter cannot distinguish between what is music and what is added noise. Hence, the noise is treated as though it was part of the music and is incorporated in the interpolated sinusoid. But the added noise is not sinusoidal and is not part of the music. It was added to mask the quantization noise in the digital data recorded on the CD. Ideally, the dither should be removed prior to upsampling and than reapplied; otherwise it becomes just more signal-correlated noise. And how to you mask signal-correlated noise? You add random noise and unadulterated jitter is a pretty good source of Gaussian noise.

I think music would be better served if digital recordings were not dithered. Dither, if it is used, must be added as the last step before the D/A conversion. The current practice of upsampling a dithered signal is absurd. If adding dither were left to the DAC, users could add as much or as little as they preferred and there would be an added bonus: Each playing of a recording would be unique and different from all other playings because a different random dither value would be added to each sample for each playing.

[JohnW]

I don’t believe there’s been any argument about lower levels of random (non-correlated) jitter being a good thing; unfortunately 99% of clock jitter is correlated – which is a REAL bad thing!

It’s always been important to differentiate between NON-Correlated Jitter and Correlated Jitter.

For a long time (since Pink Triangle Dacapo) I've added Random phase noise to my clock sources to decorrelate any artifacts, just avoided talking about it to save myself a headache....

[Guido Tent]

Originally posted by JohnW
I don’t believe there’s been any argument about lower levels of random (non-correlated) jitter being a good thing; unfortunately 99% of clock jitter is correlated – which is a REAL bad thing!

It’s always been important to differentiate between NON-Correlated Jitter and Correlated Jitter.

For a long time (since Pink Triangle Dacapo) I've added Random phase noise to my clock sources to decorrelate any artifacts, just avoided talking about it to save myself a headache....

Hi,

John is right, except for the random noise masking an intrinsically bad oscillator - throw that one out, use a real good osc (and clean up the rest of the circuit) and you'll notice there is no need for noise anyhow.

best

[Bruno Putzeys]

Mr. Ulas,

In response to an earlier post made by yourself ( http://www.diyaudio.com/forums/showt...828#post732828 ), the difference between noise and jitter (aka amplitude noise and phase noise) has been thoroughly explained. Why proceed to make the same inane statements again? Would you please read up on sampling and quantisation theory (and I mean the ones with the formulae) before throwing your weight about?

Now about dither. I understand from your profile that you know how to program. Do yourself a favor and write a program to make an FFT of 1) a sine wave fed through a quantiser and 2) a sine wave quantised after first adding rectangular probability density noise with a peak-to-peak amplitude equal to the quantiser step size. Plot 1 will show clear distortion caused by the quantisation. Plot 2 will show you that the quantisation distortion is not just masked. It's gone completely. The bin width of the FFT (you know what I'm talking about, do you?) will allow you to see distortion products well below the integrated noise floor, so you'll see that it's not just noise painting over the distortion.

Adding the right amount of noise before quantising has the effect of quantising signals between two quanta alternately to one and the other level with a statistical probability determined by how close the signal is to either. That's how it works. Masking has not a yip to do with it.

Now your absolutely smashing brilliant ingenious idea to add the noise *after* the quantisation (ie. before D/A conversion) only adds noise, but does NOT remove the distortion, because that was caused at the quantisation stage, and the noise is no longer able to influence the quantiser's decision. You will find you need to add a fantastic amount of noise before it is capable of masking the quantisation distortion.

About random and pseudorandom sequences. When writing a dither algorithm, care is taken that the pseudorandom sequence is long. The off-the shelf random number generators usually repeat themselves only once in a few hundred years, so they are quite sufficient. The ear can only tell pseudorandom noise from true random noise if it can hear the sequence at least twice within a few minutes.
The "coefficients" in pseudorandom number generators are carefully chosen primes (one needs some maths for that). Fiddling with them will result in very short sequences. This could explain why you get different colourations in your experiments. That is, if you were indeed toying with the noise generator. Actually I think you were using third party software producing noise shaped dither with user settable zeros. In which case you were manipulating the noise transfer function of the noise shaper, not the random number generator (which, as explained above, does not need or tolerate tinkering).

I'll spare you the details about noise shapers. Actually I'm afraid I'm totally wasting my time here anyhow. Writing an explanation requires second-guessing what knowledge is already present in the other party. The complete jumble of concepts as appears in your post, makes it impossible to get a feel for what you understand and what you don't. It is possible to put relative quantities on those though.

But, when you do write again, could you add something with the buzz word "noise shaper" in it? It would be hugely entertaining.

[Jocko Homo]

You are bieng kind, Bruno. If you want to say what is really on your mnd, let me do it for you. I can stand a trip to "The Cooler".

Jocko

[Ulas]

Mr. Putzeys,
Why are you so upset? Didn’t you read the disclaimer in my signature line? I enjoy tweaking the self-righteous experts in this forum.

All your equations signify nothing. Your rant echoes what I heard 30 years ago when digital audio was introduced. At that time, digital audio was PROVED to be Perfect Sound Forever. Since then I have seen a parade of explanations that PROVE how each new generation is more PERFECT than the preceding one. <<Yawn>>

I don’t listen to equations; I listen to music. After all the manipulations of the digital signal are done, it really doesn’t matter if the result has no measurable harmonic distortion. If what comes out of the D/A is not identical to what went into the A/D, the difference is noise and that’s what I hear when I listen to digital audio.

[carlosfm]

Originally posted by Ulas
If what comes out of the D/A is not identical to what went into the A/D, the difference is noise and that’s what I hear when I listen to digital audio.

That's curious.
You may try to switch on the Dolby S and tune the bias for that tape.

[Bruno Putzeys]

I think one reason why digital audio has been so far from "perfect" all these years, is because most people involved in the trade (nearly as much in the pro end as in the consumer end) worked with nonsensical hunches and assumptions in ways similar to Mr. Ulas, without bothering much about the math and the formulae, precisely like Mr. Ulas is doing. As much as he is ranting about the mediocrity of the vast majority of digital converters, he is not going to be one of the people changing the situation.

Nevertheless, a few converters have been made by people who *did* care to look up and understand the mathematical background. Needless to say, these boxes are astoundingly transparent and it takes exceedingly good ears to tell the difference between the incoming feed to the AD and the signal coming out of the DA. Unfortunately, since the designers took no shortcuts, these converters are very expensive. I strongly doubt Mr. Ulas has one of them, otherwise he would be sparing us his "thoughts".

[Jocko Homo]

Only 2 types of people believe the "perfect forever" hype the marketing weenies thought up:

The marketing weenies and people who buy their system at Best Buys.

The main problem with digital is the companies that invented the format that we are swtuck with did not care enough to make it right. It is only designed to be consumer-grade. Where "good enough" is more than good enough. Read all the minutes if the CD consortium meetings in some old JAES if you think elsewise.

I think perhaps someone here has wandered too close to Roswell.

Jocko

[Bruno Putzeys]

Indeed the slogan "perfect sound forever" was an incorrect appraisal of how digital audio works. It simply ignored that the signal had to pass A/D and D/A stages, which were anything but perfect. "The same sound forever" would've been closer to reality, though from a practical perspective not very true either.

Maybe the slogan should've been "The same sound forever, provided the same disc and CD player are used every time".

I'm not sure if the format would've gone anywhere with that slogan.