Arabic pronunciation system based on padding pre-processing and deep learning techniques

Problems in learning due to the Covid-19 pandemic have occurred in several activities e.g. teaching in Taman Pendidikan Al-Qur’an (TPA). In carrying out its activities, TPA relies heavily on the teacher to make a learning pattern on how to pronounce the Arabic alphabet of 28 letters adequately. It requires a unique approach due to the various types of sound pronunciation in reading the Qur’an. In view of health protocol rules by the World Health Organization, face-to-face meetings are replaced with online sessions, thereby affecting the learning quality. To solve this problem, a system has been developed which consists of a deep learning model. Voice data were collected for TPA students and pre-processed before the voice data were used to develop the deep learning model. Four techniques were tested to identify the best technique for pre-processing the voice data. The techniques were Spectrogram, Padding, Mel-Spectrogram, and Mel-Frequency Cepstral Coefficient techniques. The pre-processing process is to prepare the training and testing data for the deep learning stage. Test questions were later appeared with an application designed with Tkinter to lay out the exam questions. Once the user pronounced a letter, the voice is recorded and pre-processed to be predicted with a deep learning model. The quality and similarity of pronunciation of the letter is validated an Arabic speech grading algorithm. Results showed that the used of Padding technique in the pre-processing stage provides the best classification accuracy for the Arabic letter pronunciation.