Date : Wed, 14 Feb 1996 00:17:27 +0000 (GMT)
From : John Sullivan <js10039@...>
Subject: Re: Reading tapes
On Tue, 13 Feb 1996, Alan Hart wrote:
> Just a suggestion for reading tapes...
>
> Might it perhaps be worth doing some serious digital filtering of the
data?
> You could do an FFT on the whole sample, filter out everything above (say)
> 3kHz and below (say) 600Hz, then FFT it back.
Wrong approach. An FFT on the whole sample is prohibitively expensive.
Even in chunks, there are better ways, (and you will get boundary
condition problems with FFT/IFFT in chunks).
Use a digital filter (described in any book on digital signal processing)
- you probably only need a 3-4 stage filter for good results, which will
require about 9 multiplications per sample (easily possible in
'real-time').
> That should work, shouldn't
> it? I know you always get complications, with your data ending up being
> complex [do you just take the real part?] but it should remove a hell of a
> lot of noise. In fact, I was wondering whether some notch filtering would
> be possible - but presumably if you permit only 2400 and 1200Hz, you won't
> be able to register transitions between these frequencies. Or is that
> wrong?
You've not just got 1200/2400Hz and/or the transitions between these two
frequencies. I don't know exactly what effect FSK has on the spectrum of
the recorded signal, but you have a whole gamut of extra information,
probably in the range 0-2400Hz (assuming sinusoidal carriers) (???).
You can probably chop frequencies >2400Hz (2x baud rate, does sampling
theory back this claim up, or do I really mean 4800Hz?) as these will
have no useful information content, and this gets rid of tape hiss.
chopping under 60Hz will probably lose little information too, and
eliminates problems from mains hum, DC drift etc.
So basically we want a 60-4800Hz band-pass filter.
> Any ideas? Anyone remember their digital signal processing lectures better
> than me?
I'm a bit sketchy on the specifics, but most elementary DSP texts (you
can probably hit 3-4 of them in the local library) will describe the
algorithms quite clearly enough - experience shows that you don't have to
understand exactly how they work to get good enough results.
John
--
Go away cat, you make me smile too much.
<a href="http://callisto.girton.cam.ac.uk/users/js10039/">Me!</a>