The days of touch-tone hell are numbered: AT&T’s new natural-language-recognition system will fix everything. Aye right, as they say.
Date: Wed, 23 Jan 2002 06:50:46 -0600
From: “Webmaster” (spam-protected)
To: “Forteana” (spam-protected)
Subject: The days of touch-tone hell might be numbered
http://www.startribune.com/stories/789/1109766.html
AT&T’s new help service allows customers to speak freely
Kevin Coughlin
Newhouse News Service
Published Jan 21 2002
The days of touch-tone hell might be numbered.
AT&T gradually is rolling out a new help service to speed customers through the maddening maze of menus — “Press five for more baffling options” — that make simple calls an exercise in aggravation.
“How may I help you?” a computer now asks long-distance customers who call with problems.
And four times out of five, according to AT&T, the system actually understands them — whether they say “This charge is wrong,” or “You guys screwed up my bil l.”
Then it zips them to the right menu, or to a real person. If the machine’s not sure, it asks.
This is the new field of natural-language understanding, in which computers str ive to go beyond recognizing words, to grasp their meaning.
“Natural language, as opposed to speech recognition, is a big deal,” said Nigel Beck of IBM Voice Systems. T. Rowe Price is using an IBM system that talks wit h customers based on a 250,000-word vocabulary and 33,000 finance-related phras es, Beck said.
“Natural-language systems are right where the frontier is in call routing,” sai d James Flanagan, director of the Rutgers Center for Advanced Information Proce ssing.
‘Voice tone’ on the way
Researchers at AT&T Labs in Florham Park, N.J., say that “How may I help you?” is a precursor to “voice tone.” Before long, they say, folks will scrap telepho ne touchpads, computer keyboards and TV remotes, and simply tell their devices what to do.
Like “Star Trek.” Or the Bell System, circa 1900.
In those days, said Jay Wilpon, manager for speech processing at the labs, cust omers told an operator what they wanted.
“Now, the onus is on you to figure out how to navigate information using 12 but tons on the telephone,” Wilpon lamented.
Once the kinks are worked out, AT&T plans to retire its numbered menus. Then it might repackage “How may I help you?” for other industries. Imagine airline sc hedules, restaurant listings or weather reports without having to speak in code , as most computerized phone systems now require.
“You don’t have to say any magic words,” said Douglas Shurts, who oversees the AT&T program. “We have taught the computer to understand what customers are say ing, and how they are saying it.”
In the future, AT&T Labs aims to perfect a wireless computer tablet that respon ds to spoken or scribbled queries for directions and listings.
“How may I help you?” is possible largely thanks to faster computers and expand ing databases.
Speedy computers now test software in minutes instead of months. Vast databases “teach” the machines hundreds of thousands of words, plus expressions culled f rom thousands of actual customer calls.
Conversational difficulty
Since 1992, AT&T has used voice recognition to route billions of collect and ca lling-card calls, saving the company up to $300 million a year. Callers are pro mpted what to say, and the computer listens for those words, a technique called “word spotting.”
But following conversations is much harder for machines. They trip over “ums,” “ahems,” choppy grammar and colloquial expressions.
AT&T’s system ignores most of what’s said, listening for about 3,200 words that pertain to phone issues such as billing and directory assistance. From combina tions of these words, as learned from actual conversations, it attempts to glea n a caller’s intent.
“You’re never going to get 60 million people to talk the way you want them to t alk. For us, ‘natural’ is the way our customers talk,” said Allen Gorin, natura l-language guru at AT&T Labs.
“The crux of that is: Collect a large amount of data about what people actually say, learn what the salient phrases are and decipher the meaning of those. It’ s a mathematical and machine-learning problem,” he said.
Tests began about five years ago with 30,000 U.S. customers, and a national rol lout in selected markets began in August. The service is not yet available for business or international customers and works only from phones the network reco gnizes as AT&T residential customers.
Shurts said AT&T wants to make sure everything is up to speed. So far, so good.
Complaints about customer service are down 65 percent since the program was lau nched, Shurts said. He said it’s one-third more accurate than the old system, w here callers had to grope through as many as five automated menus and 25 choice
The new service shaves off about 30 seconds, he said, and ascertains the caller ‘s intent about 80 percent of the time.
A spokesman said this might thin the ranks of customer service representatives, through attrition, but he said the goal is to free them up for tough calls.
The program still stumbles over accents, odd expressions and background noises. Wireless and cable-TV calls were puzzling, too, at first.