FFT

[Martin Czech]

:::>- Take the phase values for one of the waveforms (i.e. the phase value
:::>in each FFT bin) and interpolate the magnitude values for each bin to
:::>the desired point in between.
:::
:::>- Do the same, but interpolate the phase values, too.
:::
:::But FFT is a transformation which obeys FFT(x+y) = FFT(x)+FFT(y)
:::and also linear FFT(c*x)=c*FFT(x). Where x,y element of C^n, c = const.
:::This would mean that all you'd get is a expensive way of mixing. Not
:::morphing. To morph two spectra one would have to interpolate along the
:::frequency axis. I.e. move a high frequency peak to a low frequency peak. To
:::do this one would have to recognize the peaks and define a way in which
:::they should move.
:::Similar to graphic morphing, where you usually set points that will be
:::transformed into each other.
:::

This is my understanding, too.

The trouble starts when you want to extract the peaks, ie. formants.
This means that the window length has to be much longer then one cycle
of the wave to be analysed, in order to get reasonable frequency resolution.
Look for simple sounds, like one guitar string struck, what an awfull
spectrum (Cool-Edit). Good for synthetic harmonic waves, but not so
good for real samples, because of time domain smearing.

m.c.