---
language:
- ar
license: apache-2.0
task_categories:
- automatic-speech-recognition
- translation
pretty_name: Egyptian Arabic
dataset_info:
  features:
  - name: audio_path
    dtype: string
  - name: sentence
    dtype: string
  - name: audio
    dtype: audio
  splits:
  - name: train
    num_bytes: 2211062502.6
    num_examples: 5100
  - name: validation
    num_bytes: 557950258.18
    num_examples: 1270
  download_size: 14910942412
  dataset_size: 2769012760.7799997
configs:
- config_name: default
  data_files:
  - split: validation
    path: data/validation-*
  - split: train
    path: data/train-*
tags:
- arabic
- ar
- egyptian
- dialect
- islam
---

# ar-eg-dataset

**40 h train / 10 h validation** of Egyptian Arabic from a single expert speaker —
**Prof. Ali Gomaa**, former Grand Mufti of Egypt (2003–2013), Professor of Islamic
Jurisprudence at Al-Azhar University and member of Al-Azhar's Council of Senior Scholars —
transcribed from his public lectures and released **with his express permission**.

- YouTube: <https://www.youtube.com/@DrAliGomaa> · Facebook: <https://www.facebook.com/DrAliGomaa>
- **Paper:** [A Quran and Hadith Speech Resource and Benchmark for Arabic ASR](https://doi.org/10.5281/zenodo.21927416) (preprint, DOI) · **Code:** [github.com/DrAliGomaa/quran-hadith-asr](https://github.com/DrAliGomaa/quran-hadith-asr)
- **Created by** [Mohamed Kotb](https://huggingface.co/DrAliGomaa) · **Under the patronage of** Prof. Ali Gomaa, former Grand Mufti of Egypt, who reviewed the handling of the religious material.

## Read this before quoting a number from it

This is a **formal scholarly register** — lectures — **not conversational street dialect**, and
it is **one speaker**. A model trained on it reaches **5.41 WER / 1.76 CER** on the held-out
`validation` split, but that measures *same-speaker, same-register adaptation*. On genuinely
unseen spontaneous Egyptian speech (lahgtna-v3, zero-shot) the same model reaches **17.20 WER /
6.32 CER**. Both numbers are real; they measure different things. Do not read 5.41 as general
Egyptian-dialect ASR.

## What it is for

Its purpose in the resource suite is to demonstrate **stable co-training**: adding this domain
to a Quran/Hadith mix moves sacred-text accuracy only 0.33 → 0.50 WER on Quran and
3.60 → 3.71 on Hadith, while cutting Egyptian error from ~12 to 5.41. Domains compose; adding
one does not cost you the others.

| Split | Rows | Hours |
|---|---|---|
| `train` | 5,100 | 40 h |
| `validation` | 1,270 | 10 h |

## Load it

```python
from datasets import load_dataset, Audio

ds = load_dataset("Dr-AliGomaa/ar-eg-dataset", split="validation")
ds = ds.cast_column("audio", Audio(sampling_rate=16000))
print(ds[0]["sentence"])
```

## Usage notes

- **Apply the same audio pipeline at inference** — see the
  [main dataset card](https://huggingface.co/datasets/Dr-AliGomaa/ar-quran-hadith14books-MSA).
- Text conventions: no *tashkīl*, numerals written as Arabic words.
- Released with the speaker's express permission as rights holder. Please keep the attribution.

## Citation

```bibtex
@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}
}
```