Dynamic Neural Network Model of Speech Perception

Abstract

Research results in neurobiology showed that the spatial organization of the somatosensory cortex, with linear or planar topology, seems to be the underlying support for the internal representation of the environment. This paper examines the feasibility of constructing self-organizing feature maps (SOFMs) suitable to model speech perception. The objective was to construct a class of dynamic SOFMs that can extract the time–amplitude and time–frequency features of the phonemes that appear in the formation of words. Two approaches are presented. One is based on constructing time-based embedding maps. The second method involved the construction of a dynamic SOFM having the Gabor transform as a transfer function. The time–frequency features of the speech sounds are revealed in the second approach. The results may be useful in applications of speech recognition.