Arabic Speech Dialect Classification using Deep Learning

Abstract

The growing use of dialect around the country has recently drawn interest from speech technology and research communities in dialect detection. This article aimed to identify Arabic speech dialects and classify them according to the country of speaking. This study presents an analysis and preprocessing system for audio inputs that express the Arabic dialects within 8 Arab dialects. The dataset contains 672 data and eight main subgroups, 84 samples for each of the eight Arabic dialects. Arabic dialect features are extracted and modeled using Convolutional Neural Network (CNN) techniques. The study shows the suitability and efficiency of the system, deep learning models are used instead of machine learning models. The overall results reveal that CNN’s implementation of our proposed system for identifying Arabic dialects reaches a degree of accuracy of 83%. This paper has proposed a system that showed its superiority in performance. The system converts the speech into images using the spectrogram feature, and CNN is used because it can extract features from images automatically. The study contributes to enhancing the classification process of Arabic speech dialects which is an essential issue as many of the studies working on Modern Standard Arabic (MSA), while the majority of Arabs speak local dialects, it is necessary to identify the dialect used by speakers in order to communicate with one another or before machine translation takes place.