README.md

---
task_categories:
- automatic-speech-recognition
language:
- ar
tags:
- quran
- coran
- everyayah
- ASR
pretty_name: Quran-Ayah-Corpus
size_categories:
- 100K<n<1M
license: apache-2.0
---
# Quran-Ayah-Corpus: A Multi-Reciter Arabic Quranic Speech Dataset

## Dataset Description:
Ayah-Corpus is a large-scale, multi-reciter Arabic speech dataset meticulously curated for Automatic Speech Recognition (ASR) tasks. It consists of high-quality audio recordings of Quranic verses (Ayahs) paired with their corresponding exact transcriptions. The audio is sourced from two primary repositories: Al-Quran.cloud and EveryAyah.com.

This dataset is specifically designed to facilitate the development of ASR models for Quranic Arabic, which features a distinct vocabulary, phonetic structure, and recitation style (Tajweed) compared to Modern Standard Arabic or colloquial dialects. All audio files have been standardized to a **16kHz** sampling rate to be compatible with most modern ASR pipelines.

## Dataset Structure:
The dataset is divided into train, validation, and test splits, ensuring a strict separation of reciters between the sets to evaluate model generalization to unseen voices.

### Data Splits:
| Split | Number of Samples | Number of Reciters |
|-------|------------------|-------------------|
| train | 230,254 | 38 |
| validation | 14,416 | 3 |
| test | 18,593 | 3 |
| **Total** | **263,263** | **44** |



📊 Analyzing 230254 durations (rows) from the Excel file...
Analyzing durations: 100%|██████████| 230254/230254 [00:00<00:00, 326999.12it/s]
========================================
🎯 RESULTS BY THRESHOLD (Excel-based) 🎯
========================================

--- Rows (files) shorter than 500 seconds ---
   📁 Number of rows: 230254
   🎯 Total duration: 3968417.59 seconds
   ⏰ Which is: 1102h 20m 17.59s

--- Rows (files) shorter than 400 seconds ---
   📁 Number of rows: 230253
   🎯 Total duration: 3967958.12 seconds
   ⏰ Which is: 1102h 12m 38.12s

--- Rows (files) shorter than 300 seconds ---
   📁 Number of rows: 230249
   🎯 Total duration: 3966514.47 seconds
   ⏰ Which is: 1101h 48m 34.47s

--- Rows (files) shorter than 200 seconds ---
   📁 Number of rows: 230219
   🎯 Total duration: 3959690.55 seconds
   ⏰ Which is: 1099h 54m 50.55s

--- Rows (files) shorter than 100 seconds ---
   📁 Number of rows: 229479
   🎯 Total duration: 3866971.27 seconds
   ⏰ Which is: 1074h 9m 31.27s

--- Rows (files) shorter than 90 seconds ---
   📁 Number of rows: 229066
   🎯 Total duration: 3827934.93 seconds
   ⏰ Which is: 1063h 18m 54.93s

--- Rows (files) shorter than 80 seconds ---
   📁 Number of rows: 228390
   🎯 Total duration: 3770780.24 seconds
   ⏰ Which is: 1047h 26m 20.24s

--- Rows (files) shorter than 70 seconds ---
   📁 Number of rows: 227276
   🎯 Total duration: 3687626.85 seconds
   ⏰ Which is: 1024h 20m 26.85s

--- Rows (files) shorter than 60 seconds ---
   📁 Number of rows: 225348
   🎯 Total duration: 3563535.47 seconds
   ⏰ Which is: 989h 52m 15.47s

--- Rows (files) shorter than 50 seconds ---
   📁 Number of rows: 221755
   🎯 Total duration: 3367880.79 seconds
   ⏰ Which is: 935h 31m 20.79s

--- Rows (files) shorter than 40 seconds ---
   📁 Number of rows: 215071
   🎯 Total duration: 3071315.05 seconds
   ⏰ Which is: 853h 8m 35.05s

--- Rows (files) shorter than 30 seconds ---
   📁 Number of rows: 199514
   🎯 Total duration: 2538879.34 seconds
   ⏰ Which is: 705h 14m 39.34s

--- Rows (files) shorter than 20 seconds ---
   📁 Number of rows: 163816
   🎯 Total duration: 1669293.66 seconds
   ⏰ Which is: 463h 41m 33.66s

--- Rows (files) shorter than 10 seconds ---
   📁 Number of rows: 85129
   🎯 Total duration: 535014.06 seconds
   ⏰ Which is: 148h 36m 54.06s

--- Rows (files) shorter than 5 seconds ---
   📁 Number of rows: 26516
   🎯 Total duration: 101191.18 seconds
   ⏰ Which is: 28h 6m 31.18s

========================================

### Data Fields:
Each instance in the dataset consists of the following fields:

- **audio**: A dictionary containing the raw audio data (bytes) and its sampling rate
- **duration**: The duration of the audio file in seconds (float)
- **text**: The ground-truth transcription of the Quranic verse (string)
- **reciter**: The name of the reciter (Qari) (string)

### Data Instance Example:
```python
{
  "audio": {
    "path": null,
    "bytes": "...",
    "sampling_rate": 16000
  },
  "duration": 12.17,
  "text": "الْحَمْدُ لِلَّهِ رَبِّ الْعَالَمِينَ",
  "reciter": "Karim Mansoori"
} 

```
How to Use
The dataset can be easily loaded using the datasets library:

```python
from datasets import load_dataset
# Load the dataset
dataset = load_dataset("rabah2026/Quran-Ayah-Corpus")

# Accessing a split
train_dataset = dataset["train"]

# Printing the first example
print(train_dataset[0])

```

## Reciters in each Split:
To ensure robustness and prevent data leakage, reciters are exclusively assigned to one split.

### Train Set Reciters (38):
- Abdul Basit
- Abdullah Basfar
- Abdurrahmaan As-Sudais
- Abdul Samad
- Abu Bakr Ash-Shaatree
- Ahmed ibn Ali al-Ajamy
- Alafasy
- Hani Rifai
- Husary
- Husary (Mujawwad)
- Hudhaify
- Ibrahim Akhdar
- Maher Al Muaiqly
- Minshawi
- Minshawy (Mujawwad)
- Muhammad Ayyoub
- Muhammad Jibreel
- Parhizgar
- Ayman Sowaid
- Abdullaah Awaad Al-Juhaynee
- Abdullah Matroud
- Ahmed Neana
- Akram AlAlaqimy
- Ali Hajjaj AlSuesy
- Ali Jaber
- Fares Abbad
- Ghamadi
- Khaalid Abdullaah Al-Qahtaanee
- Mohammad Al Tablaway
- Muhammad AbdulKareem
- Muhsin Al Qasim
- Nabil Rifa3i
- Nasser Alqatami
- Sahl Yassin
- Salaah AbdulRahman Bukhatir
- Salah Al Budair
- Saood Ash-Shuraym
- Yaser Salamah

### Validation Set Reciters (3):
- Mustafa Ismail
- Yasser Ad-Dussary
- Aziz Alili

### Test Set Reciters (3):
- Karim Mansoori
- Khalefa Al Tunaiji
- Mahmoud Ali Al Banna

## Dataset Creation:
The dataset was curated through a multi-step process:

- CSV files containing metadata about audio URLs and transcriptions were collected from alquran.cloud and everyayah.com
- The corresponding audio files were downloaded and converted to .wav format at a 16kHz sampling rate
- A cleaning script ensured data integrity by verifying that every audio file had a corresponding metadata entry
- A processing pipeline packaged the data with metadata into Apache Parquet files
- The data was uploaded to the Hugging Face Hub, chunked into 3 Parquet files per reciter for memory efficiency

## Licensing Information:
This dataset is licensed under the **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)** license. You are free to share and adapt the material for non-commercial purposes as long as you give appropriate credit, provide a link to the license, and distribute your contributions under the same license.

## Citation:
If you use this dataset in your research, please cite it as follows:

```bibtex
@dataset{rabah2026_ayah_corpus,
  author    = {Rabah},
  title     = {Quran-Ayah-Corpus: A Multi-Reciter Arabic Quranic Speech Dataset},
  year      = {2025},
  url       = {https://huggingface.co/datasets/rabah2026/Quran-Ayah-Corpus}
}
```

## Curated by:
This dataset was curated by Rabah.