This paper describes some experimental work in the recognition of limited-size, but arbitrary, vocabularies of spoken words. The equipment consists of a filter-bank voice-spectrum analyzer providing real-time input of measurement data to an IBM 1620-II digital computer system. The computer implements various transformations on the input data and also implements various linear decision functions which are designed by means of adaptive algorithms. Recognition experiments have investigated the recognition capability of this system on arbitrary vocabularies of up to 30 words. Several normalizing transformations on the primary measurements were investigated.
PDF