TITLE:"Fast Convolution (FFT) Filtering: From Basics to Filter Banks". DESCRIPTION: Fast Convolution filtering is a powerful technique with which every DSP engineer should be familiar. All but the shortest FIR filters are more efficiently implemented with FFTs rather than direct forms. The longer the filter; the greater the advantage. This presentation begins with the primary forms of fast convolution: overlap-add and overlap-save. These concepts lay the foundation for building parallel filters that perform mixing, filtering, and downsampling in the frequency domain. I was thinking that demonstrating the concepts via matlab/octave scripts on an LCD projector would be a good format. There will be approximately 6-10 iterations of the script, each one building on the previous concepts. I'll start by demonstrating circular convolution of two sequences using an FFT. Then add the overlapping transients of successive buffers. Then explain the computational benefit of overlap-save. ...etc In the end, the script will model a system that: performs an fft on each overlapping input block, then for each output filter/tuner: mixes via bin-shifting, multiplies by an anti-alias filter's frequency response, wraps/aliases the residual high frequency energy to the lower bins reverse ffts with a smaller transform size than the forward fft to decimate discards the wrap-around transients. I think it will fit into an hour with sufficient links for further study. It could also be easily split into two sessions. -- Mark Borgerding
comp.dsp conference: Still need presenters? How about ...
Started by ●May 7, 2004
Reply by ●May 8, 20042004-05-08
Mark Borgerding <mark@borgerding.net> wrote in news:409C4C21.9010408 @borgerding.net:> TITLE:"Fast Convolution (FFT) Filtering: From Basics to Filter Banks". > > DESCRIPTION: Fast Convolution filtering is a powerful technique with > which every DSP engineer should be familiar. All but the shortest FIR > filters are more efficiently implemented with FFTs rather than direct > forms. The longer the filter; the greater the advantage. This > presentation begins with the primary forms of fast convolution: > overlap-add and overlap-save. These concepts lay the foundation for > building parallel filters that perform mixing, filtering, and > downsampling in the frequency domain. > > > > I was thinking that demonstrating the concepts via matlab/octavescripts> on an LCD projector would be a good format. > > There will be approximately 6-10 iterations of the script, each one > building on the previous concepts. > > I'll start by demonstrating circular convolution of two sequences using > an FFT. > Then add the overlapping transients of successive buffers. > Then explain the computational benefit of overlap-save. > ...etc > In the end, the script will model a system that: > performs an fft on each overlapping input block, > then for each output filter/tuner: > mixes via bin-shifting, > multiplies by an anti-alias filter's frequency response, > wraps/aliases the residual high frequency energy to the lower bins > reverse ffts with a smaller transform size than the forward fft to > decimate > discards the wrap-around transients. > > > I think it will fit into an hour with sufficient links for furtherstudy.> It could also be easily split into two sessions. > > -- Mark Borgerding >Wonderful! Lori Ann or Rick will get you on the schedule. -- Al Clark Danville Signal Processing, Inc. -------------------------------------------------------------------- comp.dsp conference July 28 - Aug 1, 2004 details at http://www.danvillesignal.com/index.php?id=compdsp email: compdsp@danvillesignal.com Who says you can't teach an old dog a new DSP trick?