Predicting Quranic Audio Clips Reciters Using Classical Machine Learning Algorithms: A Comparative Study

Abstract

This paper introduces a comparative analysis for a supervised classification system of Quranic audio clips of several reciters. Other than identifying the reciter or the closest reciter to an input audio clip, the study objective is to evaluate and compare different classifiers performing the stated recognition. With the widespread of multimedia capable devices with accessible media streams, several reciters became more popular than others for their distinct reciting style. It is quite common to find people who recite Quran in mimicry tone for popular reciters. Towards the achievement of a practical classifier system, a representative dataset of audio clips were constructed for seven popular reciters from Saudi Arabia. Key features were extracted from the audio clips, and different perceptual features such as pitch and tempo based features, short time energy were chosen. A combination of perceptual features were also completed in order to achieve better classification. The dataset was split into training and testing sets ($80\%$ and $20\%$, respectively). The classifier is implemented using several classifiers (SVM, SVM-Linear SVM-RBF, Logistic Regression, Decision Tree, Random Forest, Ensemble AdaBoost, and eXtreme Gradient Boosting. A cross comparative results for all acoustic features and top six subset are discussed for the selected classifiers, followed by fine-tuned parameters from classifiers defaults to optimize results. Finally we conclude with the results that suggest high accuracy performance for the selected classifiers averaging above $90\%$ and an outstanding performance for XGBoosting reaching an accuracy rate above $93\%$.