Phoneme duration scheme for tajweed medd rules recognition in qur’an recitation

Abstract

The speech recognition techniques can be used to implement a Computer-Aided Pronunciation Learning (CAPL) System, however, the Computer-Aided Holy Qur’an Learning still needs more research because of the great difference between Qur’an recitation and normal speech. The major difference is due to the Tajweed rules that control the Qur’an recitation, especially those rules that depend on the phoneme duration like the Medd rules. Current speech recognition applications recognize the phonemes regardless of their duration and are not sufficient to recognize the Quranic recitation. There are two stages to get acquainted with the Medd rules: classification and estimation of duration. The previous studies classified phonemes in Qur’anic recitation, such as the classification of Arabic speech with no concern the Tajweed rules' impact, especially the classification of vowels governed by Medd rules. Regarding phoneme duration estimation, previous studies suggested a specific range for each long vowel calculated in milliseconds, ignoring the difference in recitation speeds from one reciter to another. Neglecting Medd rules in classification and duration estimation stages leads to a lack of proper recognition of Medd in the Qur’anic recitation. In this thesis, a new phoneme duration scheme is proposed to enhance the Medd duration recognition based on speech recognition techniques. A standard Qur’an recitation corpus has been collected to be used in training and testing of Medd duration. 21 Qur’an verses were chosen to cover all types of Medd, 100 famous Qur’an reciters' recitations have been collected from Web and 30 reciters have been asked to record their recitations four times for each verse at different speeds. This corpus was used to develop a Hidden Markov Model (HMM) model to recognize the Qur’anic recitation. A rule-based phoneme duration algorithm for Medd classification (RPDMCA) classifies all phonemes based on their duration and adds the required duration to each phoneme in triphone tree, thus determining the required duration for each phoneme according to Tajweed rules. In addition, an Artificial Neural Network-based Medd duration model was proposed to estimate the actual duration of phonemes. Moreover, a phoneme alignment algorithm based on the Qur'an acoustic model was developed, and the recitation rate was calculated to be used as input of the Artificial Neural Network (ANN) model. The results obtained demonstrated the high efficiency of the proposed scheme to recognize Medd types correctly. The accuracy of the phoneme classification algorithm was high ranging from 98% to 100% depending on the type of Medd. The proposed algorithm for phoneme alignment based on the Qur’an phoneme model gives a significant improvement in phoneme segmentation compared to the existing HMM Toolkit (HTK) forced alignment algorithm, where 30% of phonemes have time-error less than 30 milliseconds with manual segmentation reference. Likewise, for the Medd estimation model, it achieved results that significantly outperform the previous techniques, as its accuracy reached 86% when using manual segmentation and 70% when using automatic segmentation.