Improved Quranic ASR with Fine-Tuned Transformer Models

Automatic Speech Recognition (ASR) for Quranic recitation faces challenges with high error rates on user-recited verses and incomplete coverage of the Quranic corpus. This research presents a comprehensive empirical study focused on fine-tuning pretrained Transformer-based models to improve Quranic ASR performance. The study investigates advanced speech feature extraction methods, specifically Wav2Vec2.0, HuBERT, and XLS-R, which utilize self-supervised learning to capture context-aware speech features. These pretrained models were fine-tuned using a substantial dataset of over 870 hours of both professional and user Quranic recitations. Through detailed ablation studies, the researchers analyzed the impact of various factors, including feature extractors, output label formats, training strategies, and audio clip durations, on transcription accuracy. The most effective configuration achieved a Word Error Rate (WER) of 0.08 on the EveryAyah subset and 0.11 on a combined dataset, representing a notable five-percentage-point improvement over the Citrinet baseline. Furthermore, this optimized approach reduced training time from 140 hours to 40 hours. Key findings indicate that Arabic text without diacritics yields the best fine-tuning results, and Wav2Vec2-XLSR-53 provides the strongest overall speech representation.