Applying neural network to robust keyword spotting in speech recognition application

A word spotting recognition system is developed using an artificial neural network based on the Quickprop algorithm to recognize the keyword “collect” corrupted by white Gaussian noise in continuous speech. The neural network is constructed with 420 input nodes, 70 hidden neurons and 1 output neuron. A sigmoid function is used for activation function. Two phrases, one with the keyword and the other without are used for training. Ten phrases are used for testing on the trained network in which five versions are associated with each phrase. Misclassification happens to the original version of one phrase containing the keyword and false alarms happen to two phrases without the keyword