In this MLPs based phoneme classifier, 39 dimensional of MFCC feature vectors are served as an input data of MLP input layer which is fully-connected to one hidden layer. The output layer has 53 neurons corresponding to the Thai phonetic units. In the hidden layer, the sigmoid function is used as an activation function. A number of hidden nodes are changed through the experiments for the best results.
As all HMMs use the same MLP as source for their PDFs in continuous HMM, The output probability of each state of the HMM is computed by a weighted sum of a fixed number I of PDFs: