Designing a System to Recognize Main Arabic Dialects

Abstract

Identifying dialects can be considered as one of the recently attracted researchers. This study concentrates on recognizing between famous Arabic dialects based on speeches or talks. Three types of Arabic talks are considered here namely Arabian Peninsula Dialect (APD), Egyptian Dialect (ED), and Levantine Dialect (LD). We propose three artificial neural network models to identify between the three types of talking. These models are based on the Multi-Layer Perceptron (MLP) network, Convolutional Neural Network (CNN), and Deep Recurrent Neural Network (DRNN). Furthermore, comparisons between the three proposed models are provided. So, a comprehensive study is presented in this paper. Spoken Arabic Regional Archive (SARA) dataset is employed. It is prepared and divided into three groups. These are the Original SARA (OSARA), Filtered SARA (FSARA), and Mixed SARA (MSARA), which consists of both the OSARA and FSARA. According to the results, it has been found that the best accuracy of 90.70% is for the proposed DRNN model with the MSARA group of the employed dataset.

Keywords

• Arabic dialects
• convolutional neural network
• deep recurrent neural network
• multi-layer perceptron