[sdiy] Re: Talking about vocal filters....

[jbv]

Martin Czech :

> It always seems a total waste to me to have 20+ synthesis channels
> for a SPEECH vocoder, if some three with an additional highpass
> might do the job. I guess that the results would even be better
> since this does model the vocal tract much better then so and so many
> more poles and zeros (reverbation effect).

SNIP

> Now, is there any chance to find out the three frequencys
> in realtime from analysis input?
> Could a n-way bandpass bank do this?
> Or is a digital method the only way?

Well, that's actually what I'm planning to do with my
30 bands vocoder (someone on this list gave me the idea
a couple of weeks ago).
Since the analysis section already features a uC and
ADC + DAC (for translation of VC into MIDI ctrls and
MIDI ctrls into VC), I think I can add a peak detection algo
for the 3 main formants in input signal.
An additional board would feature 3 24dB VC-BPF, plus
a noise source + HP + VCA (for consonants, fricatives, etc).
The uC would generate from the detected peaks the requested
VC to drive the 3 BPFs, and the voice/devoice detector output
would be used to switch to noise source + HP.

The only (conceptual) problem I have so far is that in speech
(natural or synthesis) the freq of each formant evolves in time;
then, how to determine an evolving freq from peaks detected
in a sliced spectrum in the analysis section ?
In other words : when peak 1 (for instance) evolves from
band 5 to band 6, how to make sure it's still the same formant
whose freq has moved from 500 to 600 Hz...
Detecting peaks at time t and translating them into voltages is
quite easy, but tracking freqs from successive peaks in a time
sliced spectrum seems a bit more complex...

JB