Development of a Deep Learning Model for Makhraj and Tajwid Classification in Qur’anic Recitation Using CNN-Bigru and CNN-Transformer

Abstract

Accurate Qur’anic recitation depends on two complementary skills: producing each Arabic letter from its proper articulation point (makhraj) and applying tajwid rules correctly in context. Most prior audio-based studies model these two aspects as a single task or reuse the same network design for both, even though makhraj is a fine-grained phonetic discrimination problem, whereas tajwid relies on longer contextual cues. We therefore propose two task-specific deep learning pipelines. For makhraj classification, we use a CNN-BiGRU model to capture local spectral patterns and short-range temporal dynamics. For tajwid classification, we use a CNN front end followed by an attention module to model contextual dependencies. We compiled the dataset from Abdul Basit Abdus Samad’s public recitations and Ali Al-Hudhaify’s YouTube recordings, totalling about ±12 hours of raw audio before segmentation and cleaning. Three certified tajwid experts annotated the segments, with approximately 95% inter-annotator agreement. On the makhraj task, the CNN-BiGRU achieved 72.66% validation accuracy and a character error rate (CER) of 27.34%, reflecting difficulty due to phonetic similarity and class imbalance. On the tajwid task, the CNN-attention model reached 99.68% validation accuracy with a CER of 0.31% and a word error rate (WER) of 0.27%, converging in roughly 10 epochs. Overall, self-attention proved highly effective for contextdependent tajwid recognition, while makhraj remained more challenging even with sequential modelling. The results underscore the value of designing architectures around each task's linguistic properties and support the development of more accurate and pedagogically useful Qur’anic learning systems.