A real time Mandarin speech recognition method using the classification technique of hyperrectangular composite neural networks is proposed. The method can save both the storge space and the computation time which are two important factors needed to be considered in the design of a real time speech recognition system. In the method, the storage space is reduced by grouping similar cepstrum of trained pattern into states and computing the distance between states to extract classification rules instead of using the dynamic time warping algorithm.
In this paper, an algorithm is presented for the decision making of when a state should be formed and when each states should be separated by adjusting the network parameters. The algorithm is based on fuzzy membership function for measuring minimum distance among states so that the best match between the test pattern and the reference pattern is always obtained. Also, an experiment is conducted and the performance of the method is evaluated. The result shows that using method the correct recognition rate is proportional to the increasing number of the trained patterns.