whisper-large-v3-ar-eg

openai/whisper-large-v3 fine-tuned for Quran, Hadith and Egyptian Arabic at once. Same sacred-text fidelity as the MSA model, extended to a new domain — and the point of the model is that adding that domain cost almost nothing.

• Paper: A Quran and Hadith Speech Resource and Benchmark for Arabic ASR (preprint, DOI) · Code: github.com/DrAliGomaa/quran-hadith-asr
• Created by Mohamed Kotb · Under the patronage of Prof. Ali Gomaa, former Grand Mufti of Egypt, who reviewed the handling of the religious material.
• Data: ar-quran-hadith14books-MSA
  • ar-eg-dataset

Results

Normalized WER/CER (%), scored with eval/asr_score.py.

Evaluation	WER	CER
Quran (ʿAbd al-Bāsiṭ, 7,280 clips)	0.50	0.14
Hadith (Bukhari + Muslim, 4,752 clips)	3.71	1.08
Egyptian — ar-eg-dataset validation, same speaker/register	5.41	1.76
Egyptian — lahgtna-v3, zero-shot, unseen speakers, spontaneous	17.20	6.32
MGB-2	— (in this model's training set; not evaluated)

Adding a domain did not erode sacred text. Against the MSA model: Quran 0.33 → 0.50, Hadith 3.60 → 3.71 — essentially flat — while Egyptian dropped from ~12 to 5.41. That is the paper's central claim, and it is why this resource can be added to an existing mix rather than replacing it.

Be careful with 5.41. The Egyptian training data is a single speaker in a formal scholarly register. 5.41 measures same-speaker, same-register adaptation. On unseen spontaneous Egyptian we reach 17.20 and trail a general-purpose baseline (11.95). Both numbers are honest; they measure different things.

Usage

import torch
from transformers import pipeline

pipe = pipeline(
    "automatic-speech-recognition",
    model="Dr-AliGomaa/whisper-large-v3-ar-eg",
    torch_dtype=torch.float16,
    device="cuda:0",
    chunk_length_s=30,
)

gen = {
    "language": "arabic",
    "task": "transcribe",
    "num_beams": 5,
    "temperature": (0.2),
    "condition_on_prev_tokens": False,
    "compression_ratio_threshold": 1.35,
    "logprob_threshold": -1.0,
    "return_timestamps": False,
    "max_new_tokens": 444,
}

print(pipe("audio.mp3", generate_kwargs=gen)["text"])

This model does not emit timestamps. For identical accuracy with time alignment, use whisper-large-v3-ar-eg-timestamps.

Two things that will otherwise cost you accuracy

1. Run the audio pipeline first — silence-trim, loudness-normalize to −16 LUFS, edge-pad 100 ms, gain-correct, segment ≤ 30 s. Required to reproduce these numbers. pipeline/.

2. Score with eval/asr_score.py — Arabic WER moves materially with the normalizer.

Training

Base	openai/whisper-large-v3
Trained on	the MSA mix + 40 h Egyptian (Prof. Ali Gomaa) + MGB-2 folded in for scale + timestamped data
Held out	Quran, Hadith, Egyptian (10 h). MGB-2 is not held out for this model.
Learning rate	5e-6, linear decay with warmup
Batch	4 per device × 8 GPUs = 32 effective
Precision / distributed	bf16 + tf32, DeepSpeed ZeRO
Augmentation	1 clean + 4 independently-augmented waveform passes; scarce splits oversampled

Full recipe: training/training.py.

Intended use and limits

• For transcription assistance and research. Not an authority on the correct text of the Quran or hadith — verify against canonical written sources before any religious use.
• Output is Imlāʾī orthography without tashkīl; numerals are Arabic words — not for numeric transcription.
• Not a general Egyptian-dialect model. One speaker, formal register.
• Religious content reviewed and approved by Prof. Ali Gomaa, former Grand Mufti of Egypt and member of Al-Azhar's Council of Senior Scholars, under whose patronage this work was carried out; he is also the speaker of ar-eg-dataset, released with his permission.

Citation

@misc{kotb2026quranhadith,
  title  = {A Quran and Hadith Speech Resource and Benchmark for Arabic ASR,
            with Professional-Reciter Training and Validation},
  author = {Mohamed Kotb},
  year   = {2026},
  publisher = {Zenodo},
  doi    = {10.5281/zenodo.21927416},
  url    = {https://doi.org/10.5281/zenodo.21927416},
  note   = {Preprint}
}