Printer Friendly

Soliloquy for a computer's ear.

"Too bee, oar knot two B." As written this phrase doesn't make any sense, but to human ears, its sound summons up the image of a melancholy, pensive Dane. A computer "listening" to such a speech, however, would have difficlulty deciding how to spell out Hamlet's question.

This is one of many problems that automatic speech recognition systems ought to be able to solve but usually stumble over (SN: 11/16/85, p. 313). So far, most such systems recognize only a few words, spoken one at a time, often make mistakes or take an unreasonably long time to translate an incoming acoustic signal into a suitable written form.

Recently Dragon Systems, Inc., a small research and development company in Newton, Mass., demonstrated a prototype system that may bring speech recognition closer to everyday use. "It's a significant departure from anything done before," says Janet M. Baker, Dragon Systems president.

What makes the Dragon technology remarkable is the large vocabulary that it can handle--about 2,000 words--with an average response time of less than a second. Yet, combined with a simple microphone and some electronics to convert teh analog acoustic signal into digits, the system runs on a personal computer unlike other experimental systems, which often include special electronic filters and other custom-built hardware, this system does all of its speech processing in the software.

Last year, IBM Corp. in Yorktown Heights, N.Y., was the first company to demonstrate a speech recognition system able to handle sentences composed from a large vocabulary. IBM's experimental system works with 5,000 words commonly used in business. In contrast to the Dragon system, however, the computations require computer equipment worth close to $1 million. With improved hardward, IBM is now developing a 20,000-word speech recognizer.

Both the Dragon and IBM systems use "stochastic modeling." This statistical technique, for instance, allows the computer to distinguish between words that sound alike by examining the context in which they appear. By looking at surrounding words, these systems can usually handle sets of words like "to," "too" and "two."

Neither system is based on rules, which present a sequence of choices requiring yes or no answers, and neither tries to imitate the way people recognize speech. "we don't insist that certain features or characteristics of the signal must be there," says Baker. Instead, the computer estimates the probability that observed features match known language and acoustic patterns.

"Traditionally, this kind of modeling has consumed incredible amounts of computation and memory," says Baker. Dragon Systems has concentrated on finding efficiet algorithms, making it possible to run such speech recognition programs on small computers.

A person using the Dragon speech recognizer first "trains" the computer to recognize his or her speech patterns. By analyzing selected printed material, the computer also automatically builds up a model of how language is used in, say, a particular business, medical or engineering specialty.

Typically a user would speak slowly, pausing briefly between each word. The computer responds by flashing the appropriate word on a monitor, in some cases offering a menu of as many as five possibilities for anything that sounds ambiguous.

"People have responded quite favorably toward getting immediate feedback and knowing exactly what the system is doing," says Baker. By using a smaller vocabulary or a computer with a faster microprocessor, the average response time can be cut to as little as 250 milliseconds.
COPYRIGHT 1985 Science Service, Inc.
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1985, Gale Group. All rights reserved. Gale Group is a Thomson Corporation Company.

Article Details
Printer friendly Cite/link Email Feedback
Title Annotation:Dragon Systems speech recognition system
Author:Peterson, Ivars
Publication:Science News
Date:Dec 7, 1985
Words:565
Previous Article:A one-two punch for cancer.
Next Article:Bronze Age trade outpost uncovered.
Topics:


Related Articles
Legislators jump on predicted surplus.
Bed liner maker picks up.
Electronics recycling bill advances.
3 incumbents retain seats on school boards.
COMMUNITIES BRIEFLY.
PATRONS GAIN ONLINE ACCESS TO CITY LIBRARY.
'BAMAKO' PUTS HUMAN FACE ON AFRICA'S PLIGHT.
Blocking the Courthouse Door: How the Republican Party and Its Corporate Allies Are Taking Away Your Right to Sue.
Dancing from the heart: for teachers at Maryland Youth Ballet, helping students with cerebral palsy experience the joy of moving was a personal...
Making MySpace safer.

Terms of use | Copyright © 2016 Farlex, Inc. | Feedback | For webmasters