Hafs2Vec: A System for the IqraEval Arabic and Qur’anic Phoneme-level Pronunciation Assessment
Year: 2025
Type: conference-paper
Abstract: This paper details our submission-Hafs2Vecto the Iqra'Eval 2025 shared task on Arabic mispronunciation detection.Our system is built upon a wav2vec2-xls-r-1b model, enhanced by two key contributions: a strategic data mixing approach and a custom Qur'anic phonemizer.We augment the official Iqra'Eval training data with 94 hours of professional Qur'anic recitations, creating a balanced dataset that combines learner speech with high-quality acoustic references.To accurately label the reciter data, we developed a custom, Tajweed-aware phonemizer that captures the specific articulation rules of Qur'anic recitation.On the QuranMB test set, our system achieved an F1-score of 46.50% and a high recall of 79.20%.
Author: Ahmed Ibrahim
Institution: UNSW Sydney
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Funder: University of New South Wales
OpenAlex ID: W4416037791