<p>Recent advancements in Automatic Speech Recognition have significantly improved human–computer interaction, particularly for linguistically rich and diverse languages such as Arabic. However, Arabic presents unique challenges due to its complex morphology and phonetic variability. These challenges are further amplified in the domain of Qur’anic recitation, a classical form of Arabic, where the automated identification of Qira’at (recitation styles) is complicated by subtle differences in pronunciation and grammar, as well as the scarcity of annotated datasets. To overcome these limitations, this research proposes a novel model for Qira’at identification from audio recordings of the Holy Qur’an, leveraging a specialized dataset comprising approximately 1400 samples. The model architecture consists of two main components: the Visual Representation Level, which converts audio signals into Mel-Frequency Cepstral Coefficient images; and the Classification Prediction Level, which evaluates two distinct approaches: ensemble learning methods, exemplified by Stacking, Averaging, and XGBoost, and transformer-based architectures, represented by ViT, SHViT, and DeiT V3. Experimental results reveal that transformer-based models outperform ensemble learning approaches, with DeiT V3 achieving the highest performance: a testing accuracy of 99.6%, sensitivity of 99.6%, specificity of 99.9%, precision of 99.7%, an F1-score of 99.5%, and a false positive rate of just 0.12%. These findings underscore the effectiveness of transformer architectures in addressing the intricate challenges of Qira’at identification from audio-derived visual features.</p>