posted 10 November 2001 18:02            

fft_wierdness.zip

It is an FFT analysis that performs a feedback delay on the real and imaginary components. At delays equal to NFFT/2 samples the result is to repeatedly add the same spectral components over and over in the delay line. At high enough feedback levels (e.g., 0.9 or more) this tends to enhance spectral components of the original sound that are in-phase with the system sample rate divided by the FFT length.

The result produces enhancement of overtones in the input sound that match that criteria. Included is a short sample file of a piano with an arpegiator in the background. I played particular notes that seemed to "resonate" with this feedback system. Almost horn-like and bell-like at the same time.

Just fooling around... Try changing the delays to (NFFT/2-1) or (NFFT/2+1) and you get some very strange sounding sweeps as the spectral components of the previous intervals get mixed with different components in the current interval, ad infinitum...

posted 12 November 2001 21:33            

Looks like our discussions a while back sparked some interest in the FFT experiments - I'll look at your sound, and take a listen - thanks. I am in the middel of another project at the moment, so had to put the FFT stuff asside for the moment, but will come back to it in a couple of months.

posted 16 November 2001 13:38            

Cheers,

posted 17 November 2001 05:30            

Thanks for this - I'm going to be in heaven tweaking this out.

BTW, what's the purpose for the two 512 samp delays that occur
before and after the FFT/IFFT respectively?

posted 17 November 2001 11:00            

These overlapping blocks are then "windowed" on input to the FFT processing to prevent spectral leakage due to sharp edge effects of the blocks. But the window chosen can become essentially transparent to the whole process as long as you choose a "perfect mixer" window -- one of Bartlett, Hann, or Bristow-Johnson. These windows have the property that a value at one position, plus the value at a position a half-period away, sum to 1.

The output from the FFT processing must then be delayed in the opposite sense before recombining the left and right processed channels with simple mixing (addition). That realigns the data blocks before adding samples.

If you did not perform windowing then you would notice trash in the spectrum only 24 dB down or so from strong carriers. But if you only windowed the data and did not perform overlap block-processing, then you would get a serious hum in your output from the amplitude modulation of the windowing process. You need both windowing and overlap block-processing to get clean output.

[There will still be some artifacts resulting from "time aliasing" because Kyma doesn't allow multirate processing. But those artifacts should be more than 60 dB down. Multirate processing would have twice as many samples inside the FFT block processing, for the same number of samples on input. One of my earlier submissions did show a method for achieving these more pure effects, but the resulting Kyma diagrams are much more complex and the results, although cleaner, are almost indistinguishably so.]

posted 03 April 2002 09:50            

Thanks! :-)