Eyes and Ears for Computers

Abstract

Attempts to mechanize character reading and speech recognition have greatly accelerated in the past decade. This increased interest was prompted by the promise of computer inputs more flexible in format than punched cards or magnetic tape. Research has shown that automatic sensing can be done reliably if the task is suitably delimited. Cleverly designed marks on standard forms can be both machine and man readable. A single type font or a few fixed ones are tractable if the print quality is controlled. Handprinting can be handled for careful writers, as can meticulous handwriting. Isolated spoken words taken from a small number of talkers and a limited vocabulary can be automatically recognized. Typical error rates for these machine-sensings run between 0.5 and 25 per cent. These results imply that reading unrestricted typestyles, handwritten scrawl, or recognizing conversational speech is beyond the reach of present methods. From the engineering viewpoint, questions of values enter. Might it not be wiser to punch cards or tape while making copy rather than depend upon complex character recognition hardware? Is it useful to have voice input to a computer when a finger and typewriter are available? Answers to such questions will depend upon the specific application. Certainly, the utility of automatic sensing will depend upon what is to be done with the material after it enters the computer as well as the internal organization of the machine itself.