Effect of Dataset Size on Deep Learning in Voice Recognition

Abstract

Voice recognition systems mostly suffer from environmental effects and accent differences. Therefore, studies on speech recognition have begun to be examined using deep learning which is a method known to be successful in speech recognition and classification. In this study, 12 different voice commands are defined using convolutional neural network, which is a deep learning structure. In this study, the effect of dataset size on test and recognition accuracy was investigated. In addition, a different dataset which was prepared from the records of people whose main language is Turkish to investigate the effect of different accents on both test and recognition accuracy. In the experiments when the test dataset including native-speaker voice records is used, the test accuracy was obtained as 94.64% for large dataset and 64.81% for small dataset. On the other hand when the test dataset including foreigner's voice records the test accuracy reduced to 63.29% for large and 33.18% for small dataset.