Suitable CNN Weight Initialization and Activation Function for Javanese Vowels Classification

Abstract

In our previous study, a tool which employed a CNN model to classify five phonemes of Javanese mid vowel sounds dataset had been developed. Unfortunately, it’s classification accuracy was only 94%. We then tried to improve the model’s accuracy by using combinations of some weight initialization methods and activation functions. We used three weight initializations and three activation functions for the hidden layers inside our former CNN model. We also tested the dataset on logistic regression and MLP model. Experiment results show that CNN model’s accuracy with Xavier weight initialization and ReLU activation function could outperform accuracies from other models.