Getting To Know
Entrepreneur Aims To Help Computers Learn To Understand Spoken
By JOHN M. MORAN, Courant Staff Writer
Kevin Dowd is fluent in
But after years of writing computer
code and issuing arcane commands, Dowd wants to turn the
tables. It's time, he says, for computers to speak
"Humans are everywhere. Computers are
everywhere. But we're still talking different languages," Dowd
said. "The goal for me is to get something that you can talk
Enter Brainhat Corp., Dowd's new software start-up
in East Hartford, which aims to bridge the language gap
between man and machine by helping computers learn to
understand spoken English.
The launch of Brainhat comes
at a time of renewed interest in using voice to direct
computers. Faster processors, better software and the
spreading use of cellphones and other wireless devices mean
the field is ripe for an advance that would allow consumers to
get information and issue commands in conversation-like
If science-fiction writer Arthur C. Clarke had
been right, people would already be chatting away with
computers like the HAL 9000 in "2001: A Space
That hasn't happened, of course. But some
progress has been made in training computers to hear spoken
words. Voice recognition software, for example, already allows
people to dictate text and issue simple verbal commands to
But to Dowd, that merely scratches
the surface of what's possible.
"We've conquered the
menu," he said. "But this is sort of pressing buttons with
your tongue." Consequently, most PC users are still typing on
the keyboard and clicking on the mouse to get things done on
Dowd's vision is to combine voice
recognition with a kind of artificial intelligence. Together,
they would allow the computer not only to hear what's said,
but also to interpret that speech and act on it.
one thing that still isn't solved is: How do you represent
knowledge and manipulate it so that you can talk to computers
and have a conversation with them?" Dowd said.
question first fascinated Dowd two decades ago when he was
studying computer programming at the University of
Connecticut. In the years since then, he has worked for
corporate information technology departments and built his own
technology business, Atlantic Computing, which he sold two
Dowd later served as president of U.S.
operations for Articon-Integralis, a computer integration and
security firm. But he continued to pursue his interest in
natural language processing as a kind of hobby. He recently
resigned his post at Articon-Integralis to devote himself full
time to Brainhat.
Dowd has had some success building
models of his language-processing software. But don't be
looking for the HAL 9000 or "David," the robotic boy from
Steven Spielberg's movie "AI," to emerge from an East Hartford
office any time soon.
Dowd said the key to making
computers listen and understand is to sharply limit the
subject matter they must deal with at any one time. Wide-open
conversations are far too complex to tackle with today's
technology. But a simple exchange involving a single subject -
such as the weather, or movie listings, or train schedules -
is much easier to manage.
"It's a long way from talking
to your barber, but it's not so far from talking to the ticket
taker," Dowd said.
software, which is demonstrated on its website at
www.brainhat.com, allows the computer to engage in simple
conversations about the weather, restaurant food, statues and
Dowd hopes to enhance the program over time,
introducing more and more subjects until the software is able
to deal with a variety of common interactions.
Santos, an associate computer science professor at the
University of Connecticut, thinks Brainhat's idea of tackling
limited subjects is the correct approach.
take that path, it's certainly very, very feasible," Santos
said. "There's a barrier there that has nothing to do with
voice recognition and everything to do with the complexity of
making an inference about what somebody's saying."
Drew McDermott, a computer science professor at Yale
University, cautioned that natural language programming has
thus far had comparatively little success.
example, he cited an Internet search engine called
AskJeeves.com that allows users to type in English language
questions about the information they seek. "In my experience,
it didn't do any better than a keyword search," McDermott
Progress on voice recognition and artificial
intelligence tends to be incremental, McDermott said. "It's
always a big mistake to expect some big breakthough because
your expectations get pretty high," he said. "If somebody's
expecting it to behave like the little boy robot in
Spielberg's movie, that's not going to happen for a very long
Other companies are working in the field,
including large research teams at corporations such as IBM.
But Dowd said this is an area in which a small, well focused
organization can outdo a larger competitor.
has no clients or formal product lines as yet. Dowd is funding
the company, in part, with the proceeds from selling his last
company, Atlantic Computing. He expects to seek venture
capital financing as the project is refined.
computer could be made to understand English, what could it be
used for? The possibilities are seemingly
Imagine, for example, a computer that could
manage household tasks and provide information - all in
response to normal speech. What time is it? What's the weather
going to be like tomorrow? What's in the fridge? Turn down the
air conditioning a little. Set the alarm clock for 6:30 a.m.
How much heating oil is left in the tank? What was my credit
card bill last month?
Dowd thinks a variety of other
applications will emerge in time. Perhaps talking computers in
cars. Perhaps talking information kiosks in stores, public
transit stations and tourist attractions. Perhaps
entertainment options, such as a "Jeopardy"-like computer quiz
"It could be just for pure fun, if you like,"
Dowd said. "I like the idea of yelling at the house."