A neural network model for spoken word recognition

In this paper, we propose a neural approach to the speaker-independent word recognition, based on the algorithms of dynamic time warping(DTW) [8, 7] and fuzzy ARTMAP [5, 4]. DTW has some drawbacks: (1) It is space and time consuming for a large set of training patterns. (2) It gives an equal importance to each frame of a pattern. To obtain a better performance, the training patterns need to be prefiltered by human experts. Our approach attempts to address these shortcomings of DTW. We use a modified Fuzzy ARTMAP to be the framework of our approach. Our architecture is a four-layer sequential neural network. Our training algorithm and recalling algorithm are similar to fuzzy ARTMAP. However, our neural approach is a sequential algorithm. Experiments on the recognition of English alphabets have been performed. The recognition rates obtained by our approach and DTW are 87% and 80%, respectively, while memory space used in our approach is two or three times smaller than that used in DTW. Furthermore, prefiltering on training patterns is not required.