A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition

This paper studies pretrained Transformer models for Quranic automatic speech recognition, comparing Wav2Vec2.0, HuBERT, and XLS-R across speech representations, output label formats, training strategies, and dataset composition. Fine-tuned on a filtered Quranic dataset exceeding 870 hours of professional and user recitations, the best configuration achieves lower word error rates than a Citrinet baseline while reducing training time substantially. The results highlight the value of Arabic text without diacritics and identify Wav2Vec2-XLSR-53 as the strongest overall representation for this domain.