A Deep Diacritics-Based Recognition Model for Arabic Speech: Quranic Verses as Case Study

Arabic is the language of more than 422 million world’s population. Although classic Quran that 1.9 billion Muslims are required to recite, limited speech recognition exists. In Arabic, diacritics affect pronunciation a word, change in diacritic can meaning word. However, most Arabic-based models discarded diacritics. This work aims recognize while considering by converting audio signals diacritized text using Deep Neural Network (DNN)-based models. The DNN-based model recognizes DNN which outperformed traditional systems’ phonetics dependency. Three were developed speech: (i) Time Delay Network-Connectionist Temporal Classification (CTC), (ii) Recurrent (RNN)-CTC, and (iii) transformer. A 100hours dataset recordings has been used. Based on results, RNN-CTC obtained state-of-the-art results with lowest word error rate 19.43% 3.51% character rate. recognized character-by-character reliable compared transformers’ whole-sentence behaviour. performed well clear unstressed short sentences. Moreover, effectively out-of-the-dataset sounds. findings recommend continuing efforts enhancing diacritics-based obtain better performance. pretraining large could accurate recognition. outcomes be used enhance existing solutions supporting