Comparison of three back-propagation architectures for interactive animal names utterance learning

Abstract

English language is interesting for native speaker but there are many difficulties due to pronunciation. In order to facilitate for beginner to learn how to appropriately utter English word, we developed interactive learning program based on speech recognition. This paper investigates performance of three back-propagation neural network architectures with different hidden layers, e.g. 3, 4, and 5. The neural network is used to implements a speech recognition system to make interactive animal names utterance learning. The performance indicator that used in this study is number of epoch, training time, and mean square error (mse). The train dataset consist of 1, 2, and 3 syllables of animal names. The more hidden layer causes the longer training time but the smaller of the mse. Related to the number of epochs for training 1 and 2 syllables have a tendency that more hidden layers will be less the epoch, but this is not the case for training 3 syllables.