Vocoder Design questions

[Martin Czech]

> I've been considering using a single Sallen-Key bandpass (two pole) with a
> higher Q (say about 6 for 1/3 octave spacing), and an attenuator to get rid
> of the extra gain that is a byproduct (if you have access to them, the
> Electronotes Application Notes AN-25 and AN-26 have a good discussion about
> this).
> Articles I've read say the real top of the line vocoders use 5th and 6th
> order filters to get steep enough slopes to pack the passbands closer
> together.
> Is there a fault in my thinking about using higher Q two pole filters
> rather than low Q higher order filters?

Yes there is. I'm thinking about an analogue vocoder project from time
to time (sad but true, I've got almost no free time, so this is very
slow), I made some simulations (even for a whole filter bank (30) with
10 stages each,  and I've written some little c programms to compute
component values for sallen&key and multi-feed-back filters (MFB), with
given frequency and q.  These simple programms do also some variation
analysis, ie. they change all component values a little bit around the
optimum value. And this shows (as you can read in the books) that
sallen&key filters with higher q are *very* sensitive to component
value variations. This would mean to have at least 0.1% tolerance,
which is very hard to do for capacitors. MFB filters are ot so
sensitive but need on the other hand high opamp open loop gain for
higher q values in the order of >>q^2. This is a problem at the high
end of the audio band, the open loop gain of most op amps is not
sufficient. NE5532 or OP275 would be a bit better.

More stages with lower q is therefore a good idea. One could even think
about a switch to select different stages of the filter in order to
have narrow and wide bandwidth filters as desired.

And it is doubtfull if a very sharp edge bandpass design will yield a
good audio result, since narrow frequency bands or sharp slopes result
in very wide time behaviour, ie. ringing (Fourier). I think, too much q will
result in a rather reverbating output. This can be a nice effect ,
however.

If you look at the bandpass curves of the Sennheiser Vocoder you'll
find:  the frequencys are not evenly distributed (get closer together
at the high end), the same is true for bandwidth. This could be a way
to economize, since the maximum frequency resolution is in the range of
maximum resolution of the human ear. But it could be that this serves
also to equalize group delay, which seems to be important for "natural"
sounding speech (see Rabiner/Gold). If the group delay is well matched,
reverbations are minimised.

On the other hand, it is desireable to have evenly distributed filters
(a third or second away from each other), because of non-speech
applications.  Sweeping through the analyser section with a sine will
give a nice smooth glissando in the synthesizer section in this case.
Evenly distributed filters will allow nice "noise" melodies etc.  The
test CD for the MAM vocoder from Mathias Becker (German Keyboards)
shows this glissando effect (in this case not evenly distributed
filters are used, thus it sounds a bit strange, nevertheless MAM VF11
is a nice machine).

A vocoder is a major project, and there are many design constraints.
You'll learn a lot (as i did and i'll do), but it could turn out that
the sound will not satisfy your expectations. And it is a good
question, if this could be better done using dsp.

m.c.

------------------------------------------------------------------------
m.c. (alias mad scientist) has made it finally: 3 CDs à 72 minimum ready.
Visit my homepage at http://www.geocities.com/SoHo/Museum/4459/index.html