Experimental, limited vocabulary, speech recognizer

Abstract

An experimental speech recognizer is described which exhibits a high level of performance and is practical in terms of size, weight, cost, and power consumption. High recognition performance is achieved by employing the limited vocabulary approach and utilizing a simple set of parameter extractors based upon the single equivalent formant theory conceived and developed by one of the authors. Recognition logic for a vocabulary word consists of circuitry for testing the levels and movements of three parameter waveforms to determine whether they conform to the conditions of acceptability for the particular word as found from parameter data for a large number of speakers. The recognizer presently responds to the spoken digits OH through NINE with a recognition accuracy of 90 percent and an error rate of one percent on live utterances by speakers who contributed the design samples, and only slightly lower than this on other male speakers of similar speech characteristics. The recognizer occupies a volume of less than 0.8 cubic foot exclusive of microphone, indicator, and power supplies, consumes less than 30 watts, and shows promise of a very low eventual cost per word.