Abilities and limitations of a neural network model for spoken work recognition

A neural-network model for spoken-word recognition, utilizing information on loudness, is presented. The network is a discrete and simplified version of the model neural network for spatiotemporal pattern recognition presented by S. Kurogi (1987). It combines two characteristic functions: short-term storage of a kind in a synapse, and a maximum detection function. A nonnegative real number vector is regarded as the spectra pattern of a phoneme, the amplitude of a vector as loudness, and a string of vectors as a spoken work. Abilities and limitations of the network are analyzed mathematically. A time-warped word is identified as its original work if the loudness of its constituent phonemes is adjusted adequately, and successive words are segmented and recognized correctly if the loudness of the words is adjusted adequately. It is shown that the model has valid supports in physiological findings. The effectiveness of the model as an algorithm for speech recognition is also shown.<>