ART2-based multiple MLPs neural network for speaker-independent recognition of isolated words

Presents a neural network architecture for speaker-independent recognition of isolated words, which is based on the integration of multiple multi-layer perceptrons and ART2 (ART2-MLPs). In this model all the words in vocabulary are classified into several subsets according to ART2 algorithm, an original MLP is trained in order to discriminate the subsets, for the words of each subset a subset MLP is trained to recognize the words. When new classes are added, the results of previous training remains effective. Preliminary results of an experiment with 30 mandarin words on a microcomputer 286 show that the ART2-MLPs model performs better than existing MLP architecture with higher recognition accuracy, smaller amount of training and less wiring complexity. The method is well-suited for other pattern recognition tasks