The Use of Multilayer Perceptrons in Isolated Word Recognition

Multilayer Perceptrons are the most widely used Artificial Neural Networks in Isolated Word Recognition. However, these networks do not adequately modelize the temporal structure of speech. In “static” classification of speech segments with Multilayer Perceptrons, the number of input nodes is fixed a priori, while the length of speech utterances is variable. In this paper, the technique called Trace Segmentation is explored in order to fit the lengths of both input layer and utterances. This technique has been studied both with the conventional Multilayer Perceptron with back propagation and with a combination of this algorithm and stochastic learning, as well as with other strategies such as Scaly Multilayer Perceptrons. Experimental results are reported, achieving performances that range from about 80% to nearly 100%, depending on the task (Spanish Digits and E-Set). These results are comparable to or higher than those obtained with Hidden Markov Models or more conventional and expensive Multilayer Perceptrons when applied to the same task.