Roanoke Times
Copyright (c) 1995, Landmark Communications, Inc.
DATE: SUNDAY, April 15, 1990 TAG: 9004110537
SECTION: BUSINESS PAGE: A12 EDITION: METRO
SOURCE: Leslie Cauley The Baltimore Sun
DATELINE: LENGTH: Long
Defined as a computer's ability to understand human conversation, speech recognition is one of many computerized tools available to businesses today for boosting efficiency and enhancing bottom-line results.
According to experts, the technology is ideally suited for the business community because of its increasing ability to recognize and automatically correct human errors - often without the assist of a person - that can exact heavy tolls from businesses in the form of lost revenue and diminished customer confidence.
Speech recognition systems can also be used to facilitate communication with the public, a valuable tool for any business whose revenue stream hinges on customer relations, industry experts say.
While the 10-year-old technology is still in its infancy by scientific standards, the technology has its early supporters in the business community. Shearson Lehman Hutton, American Express, Amtrak and Holiday Inn are among the businesses either experimenting with the technology or already using it.
Dozens of hospitals nationwide also use speech recognition systems to track patients' treatment. At one Massachusetts hospital, the speech system has been so effective in following up on patients admitted for emergency treatment that one insurance company reduced malpractice premiums for doctors using the system by 20 percent.
As these systems become more sophisticated, and more applications are developed, speech recognition will become a valuable resource for businesses of all types, as well as society at large, many experts believe.
"Speech is man's natural means of communication," said David Pallette, director of speech research at the National Institute of Standards and Technology. "It's absurd to think that man's only communications with machines is by manual means, by putting our fingers on silly keyboards. Speech recognition is the ultimate dream of communicating with machines."
In 1990, the industry ideal embodied by Hal - a system with an unlimited vocabulary that recognizes continuous speech, including slang, humor and double entendres - is still a long way from becoming reality.
While the problem of vocabulary size has almost been solved - a system produced by Dragon Systems Inc. in Newton, Mass., offers a 30,000-word vocabulary, which is more than most people need - other obstacles have proved more difficult to overcome.
The most prickly of those is getting a computer to understand continuous human speech, which tends to be inconsistent, even with a single speaker, and fraught with verbal hurdles that trip up even the most sophisticated of computers, said Larry Rabiner, director of speech research at Bell Laboratories, the research arm of American Telephone & Telegraph Co.
"In `Star Wars' - remember R2D2? - people talked to a machine just like they would a human," Rabiner said. "We're just not there yet, and the technology's not going to be here in anyone's lifetime who's alive now."
While other experts are more optimistic about when that technology will be developed, most agree that continuous speech is by far the most perplexing problem facing speech recognition researchers today.
The reason: Even with a single speaker, sounds are variable. A person who says the word "the" several times within the same sentence, for example, won't say it exactly the same twice. That can throw off the computer, which essentially builds its vocabulary by "listening" to a speaker, recording the sound and then filing it away for future reference.
Another problem is background noise, experts note. Unless a speaker is in a soundproof room, the computer has to figure out if the sound of a slamming door is a word, or just extraneous noise.
Another complicating factor is the tendency of humans to run their words together, speech researchers say. Thus, a simple question like, "What do you mean?" may come out as "Whaddya mean?," making it extremely difficult for a computer to interpret.
Throw in regional accents, slang and a few double entendres, and a computer can all but short-circuit trying to figure out what a speaker is saying, said Robert Lucky, executive director of research at Bell Labs.
"Kids learn it on their own in the school yard, but we can't explain it well enough for a computer to understand," Lucky said.
"It's a flawed technology," Rabiner agreed. "The best you're going to do is get it right 99 percent of the time."
Most speech recognition systems today require speakers to talk slowly, with short pauses between words, in a controlled environment that is relatively free of extraneous noise. Most systems are also "speaker dependent," meaning that they must be trained to recognize the voice of a speaker.
Training can be tedious, taking several hours of repeating the same word until the computer can recognize it. The process must be repeated for each speaker who wants to use the system.
Even with those limitations, speech recognition scientists have produced some bona fide success stories. By most accounts, the list is growing.
At Mercy Hospital in Springfield, Mass., for example, emergency physicians started using a speech recognition system last year to prepare reports on patients admitted to the emergency room.
Using four workstations and software supplied by Kurzweil Applied Intelligence of Waltham, Mass., emergency-room doctors dictate patient information, including symptoms, to the computer. Based on that data, the computer automatically asks questions and makes a chart from the doctor's responses.
The doctor is notified by the computer of any diagnostic shortfalls. Likewise, doctors are notified when they are treating a high-risk patient with whom they should be especially careful.
According to John Holbrook, director of emergency room services at Mercy, the system has reduced the chances for errors amid the chaos of Mercy's emergency room, where one patient walks in about every eight minutes. Up to 80 percent of all hand-written medical reports contain errors, compared to less than 5 percent for computerized reports, he said.
The result: Malpractice premiums for doctors using the system have been reduced by 20 percent and documentation for the defense of medical malpractice suits has improved dramatically.
"It's a miracle of technology, because it really does work," Holbrook said .
Shearson Lehman Hutton, the investment banking firm, is another user of speech recognition technology.
About 40 Shearson traders are currently using a speech voice recognition system on the New York Stock Exchange floor to execute trades.
To use the system, developed by Verbex Voice Systems of Edison, N.J., traders simply pick up a special phone and speak transactions as they would normally. The data immediately appears on a computer monitor for review by the trader, who completes the transaction simply by saying "Done."
The system, which was developed for use in high-noise environments, has no problem deciphering the slang or rapid-fire speech that is common on the floor, according to Verbex.
Amtrak and Holiday Inn are planning to integrate speech technology into their national reservation systems later this year, according to Bell Labs' Rabiner. Those systems will work much like existing systems, where callers hit a number on their Touch-tone phone to reach a specific department.
With the new speech system, however, callers will only have to speak the number. That should make life a little bit easier for callers with rotary phones, who now have to wait for an operator to get help, Rabiner said.
Bell Labs is currently experimenting with a speech technology for language translation, which could be a boon for multinational companies, Rabiner said. The technology, which should be commercially available within the next decade, would allow hotels, airlines and other companies with international clienteles to replace their human translators with a speech system, he said.
American Express is experimenting with a system for automating credit card validation for the 70,000 merchants who still have to call an Amex operator each time they put a charge through, Rabiner said.
The new system, which is expected to go into a trial later this year, will allow merchants to dial into a speech recognition center and simply speak a credit card number for validation. That function is now handled by about 4,000 Amex operators.
Telephone companies are also experimenting with systems to help route collect, charge and credit card calls. Some systems under development recognize key words or phrases, such as "call collect," to handle calls.
According to Rabiner, one system under development records callers' voices to allow the computer to replay the request or question, if necessary, before deciding how to handle the call.
In the event that the computer can't decide what to do, it can ask an attendant - a person - to come on the line.
by CNB