Influence of target-vector code selection on the performance of a neural-network word recognizer

Summary form only given, as follows. A report is presented on the use of nonstandard target vectors in the training of the neural network portion of an isolated-word speech recognizer. The speech time waveform is processed into a vector of 320 numbers which are used for training a perceptron-type neural network utilizing backpropagation training. Three types of target-vector codes are used in training the neural network, and it is shown that the use of nonstandard types of orthogonal training vectors can greatly reduce the training time required. It is also shown that a neural network word recognizer achieves accuracy comparable to that of a pattern matching recognizer while requiring an order of magnitude less computing power.<>