Abstract
In this article, an isolated words recognition system using a modular structured neural network and a contextual net is proposed and the performance of the proposed system is evaluated on 40 isolated Korean words. The proposed modular structured neural network consists of two kinds of neural networks—a phoneme-discriminating neural network and a phoneme group-discriminating neural network. The discriminating neural network has a partially connected perceptron-like structure with two hidden parts capturing different features of the speech data. To improve the recognition rate, a contextual net with transitional information between phonemes is used. Experiments obtained a word recognition accuracy of 95% on the test set; the improvement of phoneme recognition accuracy with the proposed method over that of a multilayered perceptron is about 3.7%. It is observed through experiments that activation of the hidden layer corresponding to each component of the input vector over time is useful for determining some phonemes, and that the contextual informations included in the given words are effective selecting appropriate words.
Get full access to this article
View all access options for this article.