vocoder

[Martin Czech]

Sorry for wasting bandwith recently, but I was
thrown out of the list and I didn't notice it
for a while.

First of all, thanx for your advice.

>Still want to build a 20pole design ?? (;->)

Very true, indeed, that's why I'm not soldering,
but thinking and simulating for a while now.
As I said, I'm just in the feasability phase of the project,
and I have to work it out.

Tolerance :

Good point.
Well, I have the measurement equipment for 0.1% C ...
HP impedance analyser/ network analyser...
This *is* horribly time consuming,
that's why an EMS is well above $14000.

So I'll certainly do some Monte Carlo run's with an
absolute maximum
of 5% tolerance for each R & C on a complete filter band,
and I'll also do simulations with more realistic op-amp models
at the high frequency end to see, if reduced gain
and phase shift will cause any problem. 
Let's see, what the phase and magnitude distribution
will look like.

Would such a bundle of simulations be accepted as
a feasabilty study ?

I have made a variation simulation on a Bessel LP
already, and (as assumed) it shows little deviation.
That is the big advantage of multi stage filters
with Bessel low Q settings : They are not very
*instable* designs in terms of R & C tolerance.
(It turns out that a higher order,low Q design
is much more stable than a low order high Q design.
The pole location get's realy important in a
2dB ripple Tcheby, because high q sections need
perfect matching with low q sections to achieve the
proper characteristic. The high Q peak is very
sharp, and easy to miss in case of tolerance.
Also the op-amp open loop gain is not so critical
in the higher order case, isn't it ?
So, maybe the Sennheiser aproach was not so stupid,
more hardware for more tolerance stability,
but same sound ?
So that's why I believe in 20th order with maybe only
10th order performance.)

I guess, that 1% R and 2% C would still be good enough,
and I know that I'll have to measure each individual
C and to parallel it with an appropriate *compensation*
C to finally get the 2% tolerance.
Furthermore I'll have all Resistors (except extrem high
and low ends) at 10K, just a varation in C.
If I could get a "belt" of resistors, less then 1%
is possible in the very sensitive frequency area
of the human ear.

Space :

Yes, you're right, the filter bank will be area consuming.
I'll try to get a single bandpass filter (10 stages)
onto one 100x160mm card. Another card goes for detector/vca.
That's 2 cards for synthesis and 1 for analysis, total 60
cards without any option like voiced/unvoiced.
But on the other hand, it's not just a vocoder, but some
assortment of multi purpose modules for my modular.
(Fixed filter bank etc.)

Power :

Low noise op-amps need lots of power. 
I guess up to 30W for the whole vocoder.
This is another point to watch out for. 
Hot design !

Slew Rate :
>There's a PWM oscillator which controls an electronic switch in each=20
>channel, so a variable <snip>

Hmm, good idea, will eliminate any offset problem that 
other solutions (ota) have. I guess that toggling frequency
is well above audio, to avoid crosstalk problems ?
Switch to 0% duty cycle will give "freeze", but for how long ?
I guess seconds or so. I thought EMS is high end, so
I expected to freeze eternaly ;>) .

Phase Cancellation :
Of course I don't calculate all component values of
the hypothetical planned 400th order system by hand.
I use a little C prg for this purpose with two parameters :
-staggering width
-number of frequency bands
The (ideal) simulations show that some setting of the first
parameter together with 60dB/oct. slope gives better
phase figures, since the magnitude of the neighbour band
has already dropped when considerable phase shift occures.
The phase characteristic of Bessel-filters is very soft,
so no steep phase drop.

>Hope you found it interesting, nevertheless.
Sure I did !

m.c.