[sdiy] jitter analysis

[Magnus Danielson]

From: Richard Wentk <richard at skydancer.com>
Subject: RE: [sdiy] jitter analysis
Date: Fri, 09 Jul 2004 14:14:07 +0100
Message-ID: <5.1.0.14.2.20040709135732.03001e38 at pop3.skydancer.com>

Richard,

> At 14:49 09/07/2004 +0200, Czech Martin wrote:
> 
> >We do not have to argue about the real limitation of the human ear.
> >At some frequency all perception will be lost. My ears can go as far
> >as 16-17 kHz, others may reach 22kHz. But not much more.
> 
> That's for steady tones though. The perception of transients is much more 
> complicated.
> 
> You're also assuming that sampling at 44.1kHz will give you perfect 
> reproduction. Of course it won't, so just from the point of view of common 
> sense it's a good idea to sample at high a rate as possible.

The whole Nyquist theorem is actually only valid for non-transient signals, it
only applies to signals of constant amplitudes of all components. Living and
listening to transients all the time, it works as a good approximation but it
is only an approximation.

> >My assumption is that the noise component which is interesting here
> >is very low in frequency, < 10Hz.
> >After all, we do not percieve those waves as noisy, but drifting.
> 
> I suspect that's likely to be wrong. This isn't just random drift we're 
> talking about. Random drift is extremely easy to simulate in digital 
> synths, and while it makes the sound more interesting it certainly doesn't 
> make it fatter. What you get instead is more of a chorus effect, which is 
> rather different.
> 
> As an interesting digital synthesis experiment you can create (in something 
> like Csound) a generic oscillator from a bank of harmonically tuned 
> bandpass filters, each channel fed by uncorrelated noise, with very fine 
> control of Q up to oscillation.
> 
> If you turn up the Q to oscillation the sound is indistinguishable from an 
> arbitrary oscillator with the same spectrum. If you back off the Q a little 
> the sound fattens up in an interesting way. That's analogous, if not 
> exactly identical, to what I suspect is happening here.

Interesting.

> What we actually perceive is fatness as a timbral quality. It's easily 
> audible in a good single oscillator design.
> 
> I'm interested because I want to know how to model that sound digitally. 
> The fact that the quality of fatness survives digital recording and 
> reproduction proves that its absence isn't inherent to 'digitalness', so 
> much as poor modelling and DSP design. If you can replay the bitstream that 
> creates the effect, you can obviously create it from scratch digitally too.

Well, there is numerous ways to produce fatness, but it still requires good
characterisation.

> >btw.: I expect that very old transistors have a lot of 1/f noise,
> >and also shot noise. New transistors have improved a lot, simply because
> >of general semiconductor advancements over the last 40 years or so.
> >This would mean that replicas with new hardware are not the same!
> 
> This wouldn't surprise me in the least.

You can however stress the transistors up to a higher noise-level. Why have
noone come up with the idea to measure the noise-level of transistors and then
chock them up to a higher level so that you have a line of transistors with the
same noise-level that way?

Oh, I think I just came up with the idea! ;O)

There is something evil about in a controlled manner stress the components to
produce more noise, but on the other hand, if one learns to control the process
then reproducability should be rewardning and trimming of noise-levels could be
avoided by producing matching components instead. With a little skill this
could be automated... who wanted a noisy transistor? >;O)

Cheers,
Magnus