Automatic Recognition of Spoken Words

Abstract

An approach to automatic word recognition based on simple segmentation rules will be described. The speech sound wave is segmented where significant changes in articulatory configuration are reflected in easily measurable acoustic features. Examples of such articulatory changes are closures or near closures of the vocal tract, nasal coupling, and changes in excitation. The segments thus obtained do not necessarily correspond to individual phonemes or to syllables. The actual recognition program consists of three main parts. Certain features of the acoustic wave are measured first and are then used to segment the utterance. Finally, the word is identified by the sequence of these segments. This program was used to recognize words spoken in isolation by a number of male and female speakers. The results will be presented and evaluated.