A General Purpose Chipset For Speech Recognition

Abstract

A high performance, flexible, low cost chipset for automatic speech recognition is presented which can be incorporated in a large range of consumer products. The chipset recognizes isolated words in speakerdependent mode. It implements a robust, modified, two stage DTW algorithm, achieving a recognition rate above 99%, when tested with a 10 words vocabulary in an office environment. This result is obtained using only two user trained templates per word. For enabling user-friendly man-machine interface, speech compression and synthesis are also performed by the chipset.