Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!husc6!cmcl2!rutgers!umd5!vrdxhq!dgis!bms-at!stuart
From: stuart@bms-at.UUCP (Stuart D. Gathman)
Newsgroups: comp.misc
Subject: Re: Commas
Message-ID: <538@bms-at.UUCP>
Date: Tue, 17-Nov-87 18:52:07 EST
Article-I.D.: bms-at.538
Posted: Tue Nov 17 18:52:07 1987
Date-Received: Sat, 21-Nov-87 00:47:28 EST
References: <10138@brl-adm.ARPA> <1623@haddock.ISC.COM>
Organization: Business Management Systems, Inc., Fairfax, VA
Lines: 15
Summary: How IBM does it

Your second guess is almost true.  IBM's voice recognition 
technique is statistical.  The machine has a large database of
phoneme/timing distributions.  Thus, the text is reproduced
with punctuation!  Unlike AI approaches, the text is not actually
"understood" by the machine.  It is simply output to match statistically
similar utterrances used to build their database.  The results are
not 100% correct, of course.  But you can bet that it gets the example
in the ad right every time!

I do not work for IBM.  I personally think that a neural network
approach is the way to go.  (Similar in concept, but vastly different
in implementation from the statistical route.)
-- 
Stuart D. Gathman	<stuart@bms-at.uucp>
			<..!{vrdxhq|dgis}!bms-at!stuart>