Syllable recognition using integrated neural networks

Abstract

Abstract This paper describes a new syllable recognition method using the integrated neural network (INN). In this method, the recognition targets are partitioned into several groups. INN consists of a control network and several subnetworks. The control network identifies to which group the input speech belongs, and the subnetworks recognize the syllables within each group. Using INN, even if the recognition scope is large, or even if there are few training samples, the network can recognize syllables with higher recognition accuracy than conventional back‐propagation networks. Furthermore, new vocabulary entries can easily be added to an INN by adding new subnetworks corresponding to the new groups. Using the grouping method based on the manner of the articulation of consonants, the recognition accuracy is 96.2 percent for INN, compared with 95.8 percent for the conventional network architecture. This higher accuracy is obtained with 40 percent lower training costs. Using the grouping method based on the hidden layer activation patterns of a network which has learned to recognize all syllables, the accuracy is 96.0 percent.