A Real-time, Small-vocabulary, Connected-Word Speech Recognition System

摘要

The mini-computer based, real-time, small-vocabulary, isolated-word, speech recognition system is based upon an adaptive theory of speech perception. This approach achieves 98% correct recognition of the ten digits spoken in isolation and with minimal speaker dependence. It was the object of this project to extend the same approach to the recognition of connected strings of words. This required the development of a procedure for segmenting the strings into words and the alteration of the recognition algorithm to compensate for errors introduced in the segmentation procedure. The system developed recognizes the ten digits plus two control words connected in strings of up to two seconds in duration. The duration of the string is limited only by the available memory, but could be extended considerably without any deterioration in word accuracy. This connected-word, speech recognition system achieves 90% word accuracy in strings of three and four words at a time, with minimal speaker dependence. For strings of two words at a time, the accuracy can be as high as 95-97%. (Author)