MOTM

To have a true vocoder you need 2 sets of bandpass filters.
Typically you have two sections in a vocoder, the first section 
performs an analysis of sorts, the second section imparts the 
characterisitics of the input of the analysis section to a second 
signal which is typically a harmonic rich source.

To do this properly you take an input signal which is usually from a 
microphone and distribute it ti a number of bandpass filters.  The 
center frequency and bandwidth of these filters varies with the 
frequencies being spaced closer together and the band width being 
narrower in the lower frequencies of the human voice.  Each bandpass 
filter feeds and envelope follower which should have its attack and 
decay times tuned to the filter section it is following.

This gives you control voltage outputs, one for each frequency band.

A harmonic rich signal of your choice is then fed to an identical 
set of bandpass filters, but this time instead of each band feeding 
an envelope follower, each band feeds into a VCA.  The amplitude of 
each VCA is controlled by the envelope followers from the analysis 
section.  For laughs and giggles some vocoders allow you to scramble 
up which follower feeds which VCA.  

The outputs of the VCA are combined and you have your vocoded signal.

Often the uppermost frequency band at 10K or so is fed by a noise 
source to simulate the sibilant sounds of human speech.  A VCA 
controlled by the 10K filter in the analysis section controls the 
amount of noise in the vocoded signal.

To patch together a vocoder using traditional synth modules using 
your example of 21 bands, you would need 21 bandpass filters and 21 
envelope followers for the analysis section, and another 21 bandpass 
filters, 21 VCAs and a noise source for the vocoding section.

This is why good vocoders are quite expensive.  

The frequencies used are not particularly critical, but they should 
be chosen to occupy the spectrum of the human voice.  Many designers 
choose to space the filters at non-harmonic intervals so as to 
minimize resonances which might occur if the input signal were to 
correspond to the center frequency of the filter and the harmonics 
of the input frequency would also correspond to the center frequency 
of other filters at harmonic intervals.

In general the greater number of filters the better the vocoding.

It is critical that the two banks of filters be tuned identically in 
order to properly reproduce the vocal formants of the input signal.