[sdiy] Re: [Fpga-synth] Hello! and an Idea!

[Magnus Danielson]

From: Nicholas Gregorich <nicksdsu at mac.com>
Subject: Re: [sdiy] Re: [Fpga-synth] Hello! and an Idea!
Date: Tue, 5 Sep 2006 17:34:24 -0700
Message-ID: <EF1A0275-7090-4FA2-A301-D50556DBA265 at mac.com>

> On Sep 5, 2006, at 1:12 PM, René Schmitz wrote:
> >
> > You have to do an "overlap and add", i.e. you compute a lot more  
> > short length FFTs, instead of a big one. Martin Czech has mentioned  
> > this technique, in conjunction with his IR related work.
> 
> Isn't that the idea behind an FFT in the first place?  Attacking a  
> big DFT (or is it DTFT?) with a number of 2 point DFTs?  Each result  
> recursively computes in a higher and higher level until you have an  
> approximation of the actual DFT.  All attempting to minimize the  
> number of multiplies needed.

Yes, that is correct. You could figure out a scheme where you process in blocks
of M samples and then bump one block in as you bump one block out and then
finalize that FFT with the M numbers of N/M point FFTs to create the full N
point FFTs. However, there will not be much of a gain since the IFFT would
still be full-blown and need suitable crossfading.

Actually, the interested party should read up on the processing in the OGG
sound-format. It does something quite similar rather than the MPEG Audio block
processing. OGG format has already proven itself in practical audio
applications.

> Or am I missing the point here?  :)

Well, just to be a bit nasty to your nice little discussion... using FFT on a
transient signal (and transient responce) is actually making a gross
oversimplification of the signal and transient responses and hence the
artifacts.

Cheers,
Magnus