slightly ot: DAT clock deviation

[Magnus Danielson]

From: Martin Czech <czech at Micronas.Com>
Subject: Re: slightly ot: DAT clock deviation
Date: Wed, 9 Aug 2000 08:41:28 +0200 (MET DST)

> Another way would be to sacrifice one stereo channel for the (reference)
> signal, ie. the pseudo random and to have only the other one for
> measurement.  This is described in "Simulation of Romm Acoustic Using
> Head Related Transfer Functions", Gerry Beauregard et al. Dartmouth
> College, 1991.
> 
> They claimed that still they weren't able to correlate repeatet
> recordings. I can not understand why, cause the reference is always at
> once recorded on the other channel...

Well, with the recording of the reference signal they are able to do the
crosscorrelation. Now, the most straigthforward way to get precission is just
to make a LONG recording. Yes, your pseudorandom noise will loop many times,
but you don't care. I have been sitting too many nigths listening to the
pulsating noise. Now, when you do such a recording, any drifts in clocks is
at least fairly limited.

While we avoided the problem of drifting clocks by taking the clock from the
DAT, we have still not avoided the problem that a drifting clock will time
compress or time expand things according to my calculation, so you be sensitive
to clock drifts but it take a diffrent form now. The problem with clockdrift
is that 100 ppm becomes LARGE in a 100 sek measurement. However, now that we
are doing cross-correlation we are really making measurements with many
diffrent time-lengths. What happends is that we would still sence the drift,
but what happends is that only long soundpaths (reflections of say 500 ms
distance, a mere 170 meter or so) would cause a major problem. Now, luckilly
things are working FOR us, since a sound must travel rather far just to drift
1 sample away, and if we now consider that it will then be dampened quite alot
then would the leakage to just the neighbor sample be very reduced just becase
the amplitude is so small! So, some knowledge of typical acoustics comes in
and save us ;D

Making multiple separate measurements just for added precission should be
possible, but just making a longer measurement is really better, since you can
better use the data you got.

> But they also describe that they had lots of trouble because of over
> or near Nyquist frequency components of the pseudo random, that upset
> the ADC.  This is a very good hint, I have to test my DAT what happens
> for -say- 100kHz sine input etc.

You DO need a steep anti-aliasing filter. You should not settle with less than
8 poles. I would consider both Butterworth, Bessel and Gauss filters to do the
job.

> If this doesn't work, I must go linear:
> 
> shift register -> filter-> linear amp->speaker

You could apply the filtering at the input, after the mic-amp and prior to the
S/H section.

> It is possible, that Beauregard tryed to determine the relative position
> of the pseudo sequence on the tape instead of correlating each single
> recording with it's own reference channel which was recorded in the same
> take. Then I'd understand their problems.

Ah, if he was runing auto-correlation you would loose the initial delay and
run into a bunch of problems. Autocorrelation is a nice thing, but I really
prefer to do propper cross-correlation when I can.

Cheers,
Magnus