Vocoder Idea

[Martin Czech]

> > I have the impression that this channel saving trick won't work.
> > The analysis filter needs some time to measure the incomming
> > spectrum correctly, and this time is inverse proportional to the
> > filter frequency (Fourier Theory). This would mean that the filter
> > has to stay very long at the lowest frequencys in order to run in,
> > so the overall analysis response will be very slow.
> > Say fmin=40Hz ->Tcycle = 25ms, if we need 3 cycles to run in this makes
> > 75 ms just for the lowest channel. Adding time for the other channels
> > will lead to s&h update intervals of > 0.2s, this seems
> > to me as a very slow response of the overall system.
> 
> That is, if you're trying to emulate normal vocoder operation.
> 
> But the idea of inserting S&Hs into the channels' VCA control signals
> may produce some interesting FX (well, I guess it won't improve speech
> quality :-)   Has anyone tried something like that?
> 
> BTW, what about completely _synthesizing_ the control signals, e.g. using
> a shepard generator and thoroughly mixing up the channel numbers...? Or
> multiple random voltages in the LFO freq range (as used in the good old
> Elektor Formant).

Certainly. I think the EMS5000 has a "freeze" feature, which means some
kind of s&h. At least there should be the possibility to adjust each
channel manually (by control voltage), even without analysis signal.
I think the Sennheiser vocoder allows for additional patchcords
to have external control voltages as synthesis input.
It was made in a way that it would fit into a Moog modular (mechanically
as well as voltage characteristic).

This makes a nice filter bank, comb filters and the like.  By clever
wiring also the analysis filter bank could be used for audio filtering,
not only analysis. This would mean to use exactly the same circuitry
for the analysis part as for the synthesis part (the analysis part has
of course some additional circuitry like envelope followers etc.) and
this could save cost as only one type of fiter pcb has to be made.

By the way: Does anybody know of other approaches to synthesis than a
multi bandpass structure ? I mean, it sounds really stupid to have 20
high q bandpass filters in order to emulate the vocal tract, which
seems to have only 3 or 4 formants. Linear predictive methods try to
find out the formants of the incoming signal and to adapt a few
variable filters (a,q, and f).
Ie. a clever analysis section may save synthesis overhead.
But I doubt if an analogue sulution would be feasible.

m.c.