An isolated word recognition system based on a low-complexity parametrization procedure

Abstract

An Isolated Word Recognition System is presented in this paper which uses a parametrization scheme based on the two-level clipped signal Autocorrelation Function. The system prototype runs on a 64 kby. partition of a general-purpose minicomputer with quite small specific hardware requirements and, for moderate sized dictionaries (≤ 40 words), gives 95-98% recognition rates with response times better than two times real time. The system uses claasical Dynamic Programming word-matching, and its main-aimed applications are the implementation of low-cost microprocessor-based speech devices and the incorporation of Isolated Word Recognizers among the software utilities of general-purpose (mini-micro) computers. The main system features are discussed and formal evaluation tests are presented.