Phonemic speech recognition system based on a neural network

A neural-network-based system for phonemic speech recognition is described. The self-organized feature map algorithm developed by T. Kohonen (1988) is used for the network model. The solution to the problem of overlap of feature vectors that correspond to speech phonemic units is proposed. This involves identification of transient segments of the speech, voiced vs. unvoiced determination, and phonemic context. The implementation takes form as a network with labels of transients and with up to two phonemic labels corresponding to stationary regions of the speech. The system is tested using a subset of an isolated-word database consisting of twenty words