Path: utzoo!utgpu!news-server.csri.toronto.edu!clyde.concordia.ca!uunet!cs.utexas.edu!tut.cis.ohio-state.edu!ucbvax!DC@max.berkeley.edu
From: dc@max.berkeley.edu
Newsgroups: comp.music
Subject: Re: Manipulating sampled sounds.
Message-ID: <00933E08.7E4E7D40@max.berkeley.edu>
Date: 18 Mar 90 23:46:17 GMT
References: <2441@rodan.acs.syr.edu>
Sender: usenet@ucbvax.BERKELEY.EDU
Reply-To: dc@max.berkeley.edu
Organization: University of California, Berkeley
Lines: 19

One approach you could take that I have not seen mentioned in the various
replies to your query in the newsgroup is the brute force method of FFTing
the entire stream, truncating the spectrum (i.e. setting everything over
5500 Hz to zero), FFTing back, and picking every fifth sample.  This is
in some sense the "best" way to do it, and the more usual FIR approach is
a decent approximation to this.  This method probably sounds intractable
but in some cases can actually be faster than the FIR approach, particularly
if you want good accuracy.  For example, if your samples are 16 bits and
you want to preserve you S/N, you need a filter that's good to 96 dB, and
if you don't want to roll off too far below the Nyquist frequency, then you
are looking at a very long FIR filter, and consequently a long time spent
in filtering.  If you can fit your whole tune on disk, there are good 
algorithms for doing the FFT, and the IEEE DSP book that someone else pointed
you toward has a canned program for doing this that I've had a lot of luck
with.  It would probably be more sensible to do it a chunk at a time, but
this would involve thinking carefully about how to best window the FFT and
how the quality of the resulting filter compares to a FIR.

Dave Cottingham