From: Bill Stewart <stewarts@ix.netcom.com>
Date: Mon, 20 Jan 1997 10:27:44 -0800 (PST)
To: pgut001@cs.auckland.ac.nz
Subject: Re: Keyword scanning/speech recognition
Message-ID: <199701201827.KAA15122@toad.com>
MIME-Version: 1.0
Content-Type: text/plain


At 03:34 AM 1/19/97, pgut001@cs.auckland.ac.nz wrote:
>I was talking to someone recently about the feasibilty of 
>keyword-scanning phone conversations.  He thought it was probably still 
>beyond the reach of current technology, I thought it wasn't 

I'd guess it's in-between, and it depends a _lot_ on how error-tolerant
you are - do you want to catch them _all_, or just reduce the
set of tapes that you'll have humans listen to later, since you're
trolling for leads but don't mind missing the occasional one?
Speaker-dependent isolated-word recognition is pretty doable.  
Speaker-independent small-vocabulary isolated-word is also pretty good 
(though the AT&T calling-card bot does a better job when I read it numbers 
than when I rapidly tell it "calling card" early in its dialog.)  
Connected-word is a lot harder.

But word-scanning isn't the only thing "They" would troll for -
it's probably easier to look for possible matches for specific voices,
since you don't have to be too accurate.  You want to catch
Big Joey's calls from Vinnie, but don't care about his wife's calls.

Some of the problems are with modeling speech well enough to 
do good algorithms; some are just getting enough horsepower to do it
in real time.  Can "They" afford enough DSP chips to do the job?
They certainly can't follow a significant fraction of domestic traffic,
and probably not even of US-to-overseas international traffic,
but they may be able to pick out calls to/from phones of usual suspects,
at least if they're mainly trying to filter for what tapes to listen to.




#			Thanks;  Bill
# Bill Stewart, +1-415-442-2215 stewarts@ix.netcom.com
# You can get PGP outside the US at ftp.ox.ac.uk/pub/crypto/pgp
#     (If this is a mailing list, please Cc: me on replies.  Thanks.)