Small Conformer-CTC Models for On-Device Quranic Recitation Recognition

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

We release two Conformer-CTC models for Quranic recitation recognition that run offline on a phone: quran-asr-small (33M parameters, 36 MB int8) and quran-asr-tiny (16M, 19 MB). Both are trained from scratch on one laptop GPU, on 1,532 h of professional and ordinary recitation in the Hafs and Warsh riwayat. On the held-out Quran-Lab benchmark, quran-asr-small reaches 4.49% WER and 2.17% CER, second of 33 systems: only a 66M retrieval-based model (2x larger) is ahead, and it outperforms models 3.5x to 236x its size. We report what mattered: checking crowd-sourced labels against the audio, fixing defects in the reference text and audio, sharing tempo augmentation within a batch (2.5x faster training), and two evaluation effects, spelling convention and repeated recitation, that dominate the remaining word errors. Models, predictions and code: https://huggingface.co/Abdou/quran-asr-small and https://huggingface.co/Abdou/quran-asr-tiny (MIT license).