ABSTRACTThe use of neural networks for the recognition of letters of the English Alphabet spoken in isolation is investigated. Two non-linear time alignment algorithms, Dynamic Time Warping and Trace Segmentation, are looked at. Their relative performances and complexity are compared to determine the more desirable. A scaly architecture neural network is used and is trained using the error back propagation algorithm. The architecture is varied by changing the number of inputs to the network and the number of hidden units in the network. The use of different transfer functions within the processing units of the network is looked at and several versions of the sigmoid function compared. Different values of learning rate are employed and the use of different target values on the output nodes is investigated. The performance of each of these networks is compared with the aim of obtaining an optimal scaly structure for the task in question. |