Quran & Hadith Arabic ASR

Speech resources and models for transcribing the two primary sources of Islam faithfully — the Quran and the Hadith — and for helping other Arabic ASR systems do the same.

When a transcript of scripture is wrong, it is not an ordinary error: chatbots, search and summarizers increasingly answer from machine transcriptions rather than from audio, so a single plausible-sounding substitution propagates as a silent corruption of the text people read and trust. The hadith — the second primary source — had no public audio ASR dataset at all before this release; every hadith dataset that exists is text only.

This is not a release about one accurate model. The benchmark and the controlled ablation exist to show, in numbers, that any Arabic ASR system can reach this level on the Quran and Hadith without getting worse at anything else — and to hand over the data and the recipe that gets it there. A few percent WER on recitation is not an unavoidable cost of doing Arabic ASR; it is a consequence of the training mix and the orthography, and both are things you control.

Start here: ar-quran-hadith14books-MSA — the main dataset, and the full write-up: benchmark design, results, the co-training ablation, and what we recommend you do.

Datasets · ar-quran-hadith14books-MSA · ar-quran-cm17-timestamps · ar-eg-dataset

Models · whisper-large-v3-ar · whisper-large-v3-ar-eg · whisper-large-v3-ar-eg-timestamps

📄 Paper (preprint, DOI): 10.5281/zenodo.21927416 💻 Code: github.com/DrAliGomaa/quran-hadith-asr

Created by Mohamed Kotb, independent researcher — author of the paper and creator of the datasets and models. Carried out under the patronage of Prof. Ali Gomaa, former Grand Mufti of Egypt and member of Al-Azhar's Council of Senior Scholars, who reviewed the handling of the religious material.

For transcription assistance and research. Not an authority on the correct text of the Quran or hadith — output must be verified against canonical written sources before any religious use.