Computer Speech Recognition | Research & Encyclopedia Articles

Computer Speech Recognition

The advent of machines capable of recognizing human speech has been anxiously anticipated for many years. Prior to 1950, experiments were based upon the idea that language could be analyzed by syntax (the formal structure of a language) and semantics (the meanings of the individual words). Early speech recognition researchers were convinced that if all the proper word meanings and rules of grammar were stored inside a computer, the machine would be able to effectively translate languages. Unfortunately, more than $20 million dollars was spent by the military and various other government agencies on this type of research--all of which failed miserably. In the 1950s the Bell System began experimenting with a system that would allow telephone numbers to be spoken into receivers instead of being dialed, but that project also failed. Many of the problems encountered involved handling variances in accent, pronunciation, background noise or even speakers with head colds, all of which easily confused the computer. It began to look as if language translation would never be possible, since interpretation is based to a great extent on understanding, and no one knew what that was, much less how to endow a computer with it. Finally, in 1950 K. H. Davis, an American scientist at Bell Telephone Laboratories built the first machine able to successfully recognize speech. His machine could distinguish ten spoken numbers from a series of acoustic signals. Later in 1959, L. S. Green, Edmund Berkely, and Calvin Gotlieb constructed something they called The Conversation Machine, which could figure out simple questions about the weather or time and surprisingly enough, make sensible answers in reply to them.

Several more sophisticated question/answer machines appeared during the early 1960s, although these also were limited in the number and types of words which could be used. During the mid-1960s the Project MAC computer research laboratory at the Massachusetts Institute of Technology, headed by Marvin Minsky (1927-), was the source of many speech recognition ideas, although none of the experiments proved highly successful. Between 1971 and 1976, an agency known as the Advanced Research Project Agency for Speech Understanding Research (ARPASUR), set aside $15 million dollars for speech recognition research. Based on these investigations, HEARSAY-II and Harpy were developed in 1976, by Raj Reddy (1937-), V. R. Lesser, and Lee Daniel Erman (1944-) at the Carnegie-Mellon University in Pittsburgh, Pennsylvania. HEARSAY-II was capable of understanding connected speech, limited by a 1,000 word vocabulary. A few years later a similar experiment known as Hwin, was developed by Woods and Wolf and their research team. Industry giants IBM and Nippon Electric Corporation and Bell Labs have also funded research projects into voice recognition. One particularly beneficial application of computer speech recognition is in the education of the handicapped as well as people unable to operate keyboards. A speech recognition system can help deaf students to speak by providing them with visual feedback when they attempt to form words. Conversely, they can receive verbal instructions from physically handicapped pupils unable to enter information on a keyboard. Speech recognition has also been used in factories to control machinery, enter data, inspect parts, and take inventory. In some hospitals, doctors and nurses wear microphones to describe their actions to a computer that interprets and logs instructions and records. As more advanced and lower-cost speech recognition chips are being developed, scores of valuable applications and uses for speech recognition systems are being realized.