Small Talk from Computers

Developing machines that will both speak and listen

"Watch out for flak ahead," snaps the copilot of the B-17 in a voice that sounds like John Wayne's. Then the plane's bombardier gives an order in a slow Southern drawl. Snips from a grainy World War II movie? Not at all. This is part of B-17 Bomber, a home video game that Mattel will start selling this summer.

Once the stuff of science-fiction thrillers, talking machines are quickly becoming a part of modern life. Since 1978, when Texas Instruments introduced the loquacious learning aid Speak & Spell, a brave new world of chatty machines that use computer chips to make them talk has been moving into factory, office and home. This year the market for talking computer products is expected to total $50 million, and by 1985 it could reach $300 million.

Machines are starting to speak up in a wide range of situations. Exit-Us of Easton, Conn., is marketing signs for hotels and other public buildings that broadcast the word exit in up to three languages during a fire or other emergency. Next year Chrysler reportedly will introduce a car that will tell the driver when the gas supply and oil pressure are low. Three weeks ago, the Federal Aviation Administration approved the Digital Readout Weather System, a machine that gives incoming pilots the time of day, temperature, barometric pressure and other vital information. Says Exit-Us President Tim English: "We will soon be talked at in ways we can't even conceive of now."

Having overcome the main obstacles to develop machines that can reproduce the human voice, researchers are now tackling the next major hurdle: manufacturing ones that will be able to listen. A number of large and small companies are working on a host of devices designed to tap the potentially huge voice-recognition market. A team of IBM researchers, for example, is laboring on the dictating machine of the 21st century. The future executive will talk into a machine that will automatically turn his spoken words into a printed text that is displayed on a video screen. With a push of a button, he could also have a copy of the memo on paper. IBM thinks it might be able to give a laboratory demonstration of the technology in about five years.

Research into voice-recognition machines is particularly active in Japan. Nissan Motor Co. last October unveiled its model of a car for the handicapped that uses voice recognition to adjust the driver's seat and outside rear-view mirror, as well as to turn on the windshield wipers, lights and radio. A group of scientists at Tokyo University has developed a machine that allows an immobilized patient to change the position of his bed and order food and drink from a robot nurse.

Though voice-recognition machines have an almost unlimited potential, experts admit that they must first overcome some serious technological problems. For one thing, the intricacies of identifying the human voice have proved to be vastly more complicated than researchers originally expected.

In order to synthesize speech, a computer must first translate a word or command that has been given to it into digital signals, which it then interprets back into sounds. Before a computer can listen, it also changes spoken words to numbers, but then it must compare the result with speech patterns already stored in its memory. Most voice recognition machines can still only recognize the voice of one person. When anyone else talks to it, the computer is incapable of understanding what is said. Even when the right person is talking to it, the machine can easily stop functioning if one of those people uses words outside its limited vocabulary or if the person's voice is distorted by a head cold or a sore throat.

Another stumbling block for machines that listen is their high cost. A computer that understands fewer than 100 words is now about $2,000, while one that recognizes 800 words can cost more than $10,000. Thus for the next few year people may have to take orders from machines without having the chance to talk back.

This file is automatically generated by a robot program, so viewer discretion is required.