Prev: 64 x 64 FFT algorithm for 4096 pt. FFT
Next: ZSR and ZIR responses....
From: VelociChicken on 30 Jun 2010 12:11
Hello, I'm doing some audio processing in the frequency domain.
I have a 2048 and a 512 FFT running parallel (overlap and add) and they have
been nulled so that the larger FFT runs from 0 to 3Khz and the smaller one
from 3Khz to 44.1Khz. SR = 44.1Khz.

When I add the output of these two, I get a dip at the frequency cross-over
point of the waveform.
This is understandable, not least because of the steep difference in the
transition bands, so I sloped off (linearly) the end of the larger FFT
magnitudes and overlapped them so the combination is nearly perfect, with an
inaudible little dip going down 5bB or so at the crossover point.

What method should I use (other than the wonderful trial and error) to mix
these signals the best way possible? Plus, for different combinations of
sizes.
I'm guessing it's something to do with mimicking the roll-off curve of the
smaller FFT, with the likes of Window-Sinc formulas?
I don't really want to up-sample the 512 FFT to a 2048 if it can be avoided.

Thanks,
Dave.


From: VelociChicken on 1 Jul 2010 09:21
"VelociChicken" <turn(a)invalid.com> wrote in message
news:A0KWn.33899$m87.9815(a)hurricane...
> Hello, I'm doing some audio processing in the frequency domain.
> I have a 2048 and a 512 FFT running parallel (overlap and add) and they
> have been nulled so that the larger FFT runs from 0 to 3Khz and the
> smaller one from 3Khz to 44.1Khz. SR = 44.1Khz.
>
> When I add the output of these two, I get a dip at the frequency
> cross-over point of the waveform.
> This is understandable, not least because of the steep difference in the
> transition bands, so I sloped off (linearly) the end of the larger FFT
> magnitudes and overlapped them so the combination is nearly perfect, with
> an inaudible little dip going down 5bB or so at the crossover point.
>
> What method should I use (other than the wonderful trial and error) to mix
> these signals the best way possible? Plus, for different combinations of
> sizes.
> I'm guessing it's something to do with mimicking the roll-off curve of the
> smaller FFT, with the likes of Window-Sinc formulas?
> I don't really want to up-sample the 512 FFT to a 2048 if it can be
> avoided.
>
> Thanks,
> Dave.

Nevermind, I've used a quater sinewave to round off the edge in 8 points,
and it's mixing with a tiny 0.25 dB error now. It's not perfect but, hey.





 | 
Pages: 1
Prev: 64 x 64 FFT algorithm for 4096 pt. FFT
Next: ZSR and ZIR responses....