Deep Acoustic Modelling for Quranic Recitation

Abstract

The Holy Quran, the only scripture of the universe preserved in its complete original text since its revelation, is of utmost importance to the Muslim community. Revealed originally in the Arabic language, we need to understand and practice how it should be recited and memorized according to the rules set out by native Arabic speakers. With the advent of AI technology in acoustic modeling, researchers began developing models of various languages; however, due to the variety of accents and dialects of Arabic, it is challenging to develop a robust acoustic model for Quranic recitation. In this research, we developed a deep learning model that is not only robust to the above linguistic properties but is not affected by the recitation styles and intricate Tajweed. When used for classification tasks, deep features from this model produced a maximum accuracy of around 96.30%. To illustrate the importance of our deep learning network as an acoustic model, a content-based verse retrieval system (CBVeRse) was developed by employing the model trained in the previous step with an Average Normalized Modified Retrieval Rank (ANMRR) of 85.39% and mean Average Precision (mAP) of 96.52%.