posted 10 July 2001 12:22            

Human speech has important amplitude changes in the 2-6 Hz band. After FFT of incoming spectra, I want to band-reject this 2-6 Hz band on each real and imaginary component. That gives me a better noise field estimate sans the speech components. The result significantly enhances the fricative and glottal sounds making it easier to understand the recovered speech.

Once the FFT has been performed, the coefficients are being updated at (2 * SignalProcessor sampleRate / ?NFFT) and so this is the internal sample rate that needs to be used for coefficient filtering. With only a single clock in the system it is difficult to see how, even with a sample-hold, to perform this downrate filtering. (...but maybe I need to sharpen my pencil before complaining too loudly...)

I have been able to do very simple minded IIR filtering with delay lines equal to (?NFFT/2) length, but now for band reject I need to perform more sophisticated IIR filtering. [...ow that I think about it, I can probably do this with multiple delays all of length (?NFFT/2)].

... also useful for very narrowband filtering where your graphic EQ sound could be used in a scaled frequency manner...

Anyway, one for the wish list...