Enhancing Quranic Recitation Accuracy Using State-of-the-Art Audio Classification Techniques

- In this study, we investigate the potential of using state-of-the-art neural network architectures to increase the accuracy in classification of Quranic recitations per verse. Using a dataset of more than 4000 audio recordings annotated with rich metadata, the research has concentrated on differentiating accurate recitations through Hukm Al-Noon Al-Mushaddah specifications. This study uses three pre-trained deep learning models (Inception-V3, EfficientNet and MobileNet), as well as hybrid model proposed in this paper to perform classification of recitations. The raw audio inputs were converted into spectrograms for feature extraction and classification in each of the models. Experiments demonstrate that through the fusion, this hybrid model significantly outperforms individual predictions by dramatically improving precision, recall and F1-scores in five different verses. The total accuracy for the proposal model is 0.79 which is the highest comparing with Inception-V3 was 0.75 and EfficientNet was 0.73. The results underline the ability of such systems to provide immediate feedback for learners and thereby assist them in adhering to traditional recitation standards, a feature that helps maintain the originality of Quranic recitation. To enable usage on real data, further work should build a bigger dataset (samples of data) and optimize the model to providing feedback with larger latency