Convolutional Neural Network for Automatic Speech Recognition of Filipino Language

Abstract

© 2020, World Academy of Research in Science and Engineering. All rights reserved. Researches focusing on deep learning for speech recognition are integral for the successful implementation of natural language processing (NLP). Successful implementations of NLP would allow even non-technical and illiterate users to have access to technology by simply using their native tongue. Currently, there are very limited studies on the use of the Convolutional Neural Network (CNN) for Automatic Speech Recognition of Filipino Language. This paper presents a CNN model using SqueezeNet architecture for the Filipino language with 99.58% training accuracy, and 84.71% testing accuracy. Experimental method was employed to achieve this accuracy by adjusting the learning rate of the model. SqueezeNet architecture was chosen because it requires less resource but maintains the same level of accuracy as compared to other neural network architectures. Mel Frequency Cepstral Coefficient was also utilized to convert the audio inputs into Mel Spectrograms. The study also presented the precision rate of each Filipino class identified in the data set and how CNN was used to increase prediction accuracy. This CNN model in this study can be used as basis for further improvement of the speech recognition rate of other Filipino words and other new languages.