A Machine Learning Approach to Arabic Phoneme Classification through Ensemble Techniques

This paper uses Machine Learning techniques to classify the pronounced Arabic phonemes into separate classes. Our object revolves around building a model that utilizes the Mel Frequency Cepstral Coefficients (MFCCs), characteristic of audio, to build an optimized classifier. For our data set, we have collected audio of Arabic phonemes from Jamia Darul Uloom Karachi. We have then pre-processed data to make it clean and equal to each other in size. To make it diverse, various techniques including Discrete Fourier Transform (DFT), mel-frequency scaling, logarithmic amplitude calculation, and the Inverse Discrete Cosine Transform (DCT) have been applied. Then individual classifiers are trained using Machine Learning techniques such as Support Vector Machines (SVM), k-Nearest Neighbors (k-NN), Decision Trees, Random Forest, and Naïve Bayes applied to the dataset for classification purposes and then are analyzed based on four metrics, namely: Accuracy, Precision, Recall, and F1-Score. SVM turned out to be the best among the others, with the highest accuracy of 97.4%. We have then used ensemble learning methods, including Bagging, Boosting, Stacking, and Voting, to enhance the classification accuracy. Ultimately, Stacking with Random Forest has better performance in all four metrics compared to other ensemble methods.