AI-Powered Qur'anic Verse Identification via Speech Recognition and Semantic Matching

Abstract

Accurate recitation of Qur'anic verses is a cornerstone of Islamic education. However, non-native Arabic speakers and learners with limited access to qualified instructors face persistent challenges in identifying and verifying recited verses. This paper presents Qiraah Finder, an end-to-end AI-powered Qur'anic verse identification system that integrates a fine-tuned Wav2Vec2 multilingual automatic speech recognition (ASR) model with a Term Frequency-Inverse Document Frequency (TF-IDF) and cosine similarity verse-matching engine. The system transcribes recited audio of Surah Al-Ikhlas, a short surah of four lexically distinct verses, and retrieves the corresponding surah and ayah from an authenticated reference corpus. Two audio preprocessing pipelines and four hyperparameter fine-tuning experiments were evaluated using Word Error Rate (WER) and Character Error Rate (CER). Baseline preprocessing substantially outperformed data augmentation (WER: 0.12% vs 3.21%; CER: 0.68% vs 2.61%), indicating that excessive augmentation is detrimental to high-quality, domain-specific Qur'anic corpora. The best configuration (Experiment B: partial fine-tuning with frozen feature-extraction layers, learning rate 3×10⁻5, batch size 4, 10 epochs) achieved the lowest WER (1.73%), a competitive CER (0.85%), and the lowest evaluation loss (0.0287). Verse matching attained 100% identification accuracy across all test samples, including a partial recitation, confirming the robustness of TF-IDF retrieval. A key limitation is that evaluation was confined to Surah Al-Ikhlas, which does not capture the full complexity of the 6,236-verse corpus. These findings establish a replicable benchmark for AI-assisted Qur'anic verse identification; future work will extend the system to longer surahs with higher lexical overlap, real-time tajweed feedback, and GPU-accelerated hyperparameter search.