FPGA Central

I recently asked about 'reconstituting' a signal which was l-o-n-g
compared to window and was lead to links which showed me the 'obvious'.
Overlap inverse of forward fft by "small" time step.
[ I'm taking *EXTREME* liberties with terminology ;]

I'm also interested in a inverse ( obverse ? ) problem.

Assume a constraint that a forward/inverse FFT be limited to nK points.

Assume also that frequencies of interest span octaves/decades.

For the sake of 'argument' assume:
signal lasts >>> hours
window = 100 msec

I understand that I can do appropriate manipulations in frequency
domain if I do the 'inverse' transform at least every 50 msec.

How to approach the problem if there is also significant data at a
rate < or <<<< than .1 second?

[ My universe of discourse favors text-to-speech and speech-recognition ;]