On Tue, Oct 18, 2011 at 11:20 AM, Roman Haefeli reduzent@gmail.com wrote:
If you're interested only in the lower frequencies anyway, you can drastically lower the sample rate in order to save some CPU cycles when doing the FFT.
Right. A 2^17 point FFT is on average not expensive at all, but the problem is that all computations (in the order of 17 * 2^17) are done at once every 3 seconds, that's just too much and it causes buffer underrun in the realtime dataflow.
Depending on input signal, drastically lowering the sample rate may also require drastic filtering to prevent aliases, which could otherwise spoil analysis results. Such filtering can be done in frequency domain, using FFT again. Be sure to use 4 times overlap and Hann windowing before FFT and after IFFT.
Katja