A Speaker-Independent Digit-Recognition System

With the widespread growth in the use of digital computers, there has been an increasing need for man to be able to communicate with machines in a manner more naturally suited to humans. The realization of this need has motivated a great deal of research in automatic recognition of speech by computer. 1-3 Although only a moderate degree of success has been obtained in solving the problems associated with machine recognition of continuous speech,4 a greater degree of success has been obtained in recognition of isolated words from a fixed vocabulary. The performance of these systems range from about 92 percent correct decisions for 561 isolated words by an individual for which the system has been carefully trained 5 to nearly error-free performance for the recognition of a limited vocabulary (e.g., the digits) also spoken by a speaker for which the system has been trained. 6 However, performance of many of these word-recognition algorithms is radically degraded when the system has not been tuned to the speech characteristics of the individual user. The subject of this paper is an isolatedword, digit-recognition system that achieves high accuracy without having to be trained every time a different speaker wishes to use the system. 81