Syllable recognition using integrated neural networks

For the purpose of syllable recognition, the integrated neural network (INN), which consists of a control network and several subnetworks, is proposed. To train INN, the recognition targets are partitioned into several groups. The control network identifies the group to which the input speech belongs, and the subnetworks recognize the syllables in each group. INN has the following advantages over a conventional backpropagation (BP) network: (1) training time is reduced by half, (2) greater recognition accuracy is obtained with fewer training samples, and (3) new vocabulary entries can be easily added to an INN by adding new groups. Two kinds of methods for partitioning syllables into groups are proposed. One is based on a priori phonological knowledge and the other on the hidden-layer-activation patterns of a network that has learned to recognize all syllables. Using INN, consonant recognition accuracies of 96.2% and 96.0% are obtained for each grouping method, respectively. A new training method that is capable of generating new data from given data is introduced. Using this method, the INN's training time is reduced by 85%. For conventional BP networks the reduction in training time is 90%.<
>