Newsgroups: comp.speech
Path: pavo.csi.cam.ac.uk!doc.ic.ac.uk!agate!howland.reston.ans.net!spool.mu.edu!sol.ctr.columbia.edu!sol.ctr.columbia.edu!mhall
From: mhall@occs.cs.oberlin.edu (Matthew Hall)
Subject: Good, simple fast filters for speech?
Message-ID: <MHALL.93Apr11144141@occs.cs.oberlin.edu>
Sender: nobody@ctr.columbia.edu
Organization: Oberlin College Computer Science
Distribution: comp.speech
Date: Sun, 11 Apr 1993 19:41:41 GMT
X-Posted-From: occs.cs.oberlin.edu
NNTP-Posting-Host: sol.ctr.columbia.edu
Lines: 34

Hello,
	I have gotten a miniature speech recognition package working
for my mac (about %80 for digits) by using a second order state space
filter (I think it's what it's called) described in the FAQ, and then
doing DTW on the patterns.  I.e
		
		|x|= r|cos(t)  sin(t)| |x+sample|
		|y|   |-sin(t) cos(t)| |y	|
Now, this works to some accuracy, and it would probably be better if I
could filter out some noise and improve my endpoint routines.
However, this filter, which I have optimized and written in machine
language FPU code still takes a fair amount of time (About 2x the time
of input at 11kHz sampling rate for the set of 10-16 frequency bands I
look at).  I have seen various sound editing programs that do fairly
high resolution spectrograms very quickly on my same machine.  It
would be very nice if someone could point me towards some info and
algorithms for fast, fast filtering or spectral analysis.  Ideally, I
would like to get the filtering to occur simultaneously with the
recording so that I could do it in the sound-driven interrupts (on a
Mac) but that means the filtering must be 4-8 times faster.  Any
source (pref. in Pascal or Mac asm) code would be appreciated.
	If I get the accuracy and speed up, and the bugs out, I will
post code if there is interest.

Thanks,
-matt


--
-------------------------------------------------------------------------------
Matt Hall.    mhall@occs.oberlin.edu
              (216)-775-6613 (That's a Cleveland Area code. Lucky Me)
  
"Life's like mayonnaise soda"	-Lou Reed
