MultiheadSelfAttention vs. Traditional Encoders: A Benchmark Study on Precision and Recall in Tajweed Recognition

Accurate classification of Tajweed rules is critical for preserving the integrity of Quranic recitation. While prior studies relied on traditional machine learning (ML) models with handcrafted features, they faced limitations in handling class imbalance and capturing nuanced acoustic patterns. This paper introduces a novel encoder-based framework for Tajweed rule classification, combining adaptive feature extraction (Mel-Frequency Cepstral Coefficients, openSMILE) with advanced neural architectures (DNN, LSTM, Self-Attention, Multihead Self-Attention). We curate a diverse dataset of Quranic recitations that span four Tajweed rules (Ikhfa, Izhar, Idgham, Iqlab) from eight reciters, ensuring robustness through varied vocal styles and regional accents. Our Multi-Head Self-Attention encoder achieves 98.6% accuracy, with perfect precision (1.00) for the minority Iqlab class, while the DNN encoder attains 98.9% accuracy—significantly outperforming prior benchmarks (82.2% accuracy with Random Forest). The LSTM encoder, however, fails to classify Iqlab (0% F1-score), highlighting the superiority of attention-based mechanisms. Results demonstrate that encoder architectures effectively address class imbalance and capture intricate temporal dependencies, establishing a new state-of-the-art for automated Tajweed validation. This work contributes (1) a preprocessing pipeline for audio standardization, (2) empirical validation of encoder models, and (3) a canonical, expert-annotated Tajweed dataset to advance research in Quranic recitation analysis.