Pengecaman sebutan digit menggunakan rangkaian neural : satu kajian terhadap carian bilangan nod. tersembunyi dan parameter pembelajaran.

Abstract

This research is about choosing suitable number of nodes in hidden layer and learning parameters in the proses of Malay number speech recognition system using neural network. The learning approach used in training Feed Forward Multilayer Perceptron (FFMLP) is back propagation. Speech for the study are analysed using Linear Predictive Coding (LPC) and Log Area Ratio (LAR) to represent speech signal for every 20ms. The neural network learning operation are greatly influenced by the parameters ie. momentum , learning rate and number of hidden nodes choosen. Choosing unsuitable parameters will make the neural network system give an answer divert from the desired one.