Design of hierarchical perceptron structures and their application to the task of isolated-word recognition

Several design strategies for feedforward networks are examined within the scope of pattern classification. Single- and two-layer perceptron models are adapted for experiments in isolated-word recognition. Direct (one-step) classification and several hierarchical (two-step) schemes have been considered. For a vocabulary of 20 English words spoken repeatedly by 11 speakers, the word classes are found to be separable by hyperplanes in the chosen feature space. Since for speaker-dependent word recognition the underlying database contains only a small training set, an automatic expansion of the training material improves the generalization properties of the networks. This method accounts for a wide variety of observable temporal structures for each word and gives a better overall estimate of the network parameters, which leads to a recognition rate of 99.5%. For speaker-independent word recognition, a hierarchical structure with pairwise training of two-class models is superior to a single uniform network (98% average recognition rate).>