| Larger window sizes will give you better low frequency resolution, but will have more variability in the high frequency range.
Small window sizes will average out the low frequency information because each window gets de-trended (has it's average, and often linear slope removed). Instead of thinking of your window size in number of points, think of it in the time domain. @ 44.1 kHz, 2048 samples is 0.046 seconds, which corresponds to a frequency of 21.5 Hz (1/0.046-21.5). This is approximately the lowest frequency you can resolve with this window size.
Another problem with smaller window sizes is that they have coarse low frequency resolution (not a whole lot of points where the FFT is calculated)...instead, most of the points are distributed in the high frequency area.
A better way to approach this would be for plug-in developers to use a Chirp-Z transform instead of FFT. It would take more processing time, but would give greatly improved low frequency accuracy. Probably this would be overkill for real-time spectral analysis plugins, but would be nice for acoustic analysis.
Cheers
Kris |