README.md

---
language:
- ar
license: cc-by-4.0
task_categories:
- automatic-speech-recognition
tags:
- quran
- arabic
- audio
- islam
- speech-recognition
pretty_name: EveryAyah Quran Recitations
size_categories:
- 10K<n<100K
---

# EveryAyah Quran Dataset

## Dataset Description

This dataset contains verse-by-verse (ayah) Quranic recitations from 4 professional reciters, with fully diacritized Arabic text aligned to each audio segment. The dataset is designed for automatic speech recognition (ASR), speaker identification, and other speech processing tasks focused on Classical Arabic and Quranic recitation.

### Dataset Summary

- **Total Audio Files**: 24,944 ayahs
- **Number of Reciters**: 4
- **Audio Format**: MP3, 64 kbps
- **Sampling Rate**: 16 kHz
- **Text**: Fully diacritized Arabic (with Tashkeel)
- **Total Duration**: Approximately 174 hours
- **Coverage**: Complete Quran (114 Surahs, 6,236 verses per reciter)

### Supported Tasks

- **Automatic Speech Recognition (ASR)**: Train and evaluate models for transcribing Quranic Arabic audio
- **Speaker Identification**: Classify reciters based on their voice characteristics
- **Speech Synthesis**: Generate natural-sounding Quranic recitations
- **Diacritization**: Develop models for automatic Arabic text diacritization
- **Pronunciation Analysis**: Study Quranic recitation rules (Tajweed)

### Languages

- Arabic (Classical/Quranic Arabic)

## Dataset Structure

### Data Instances

Each data instance contains the following fields:

```python
{
    'audio': {
        'path': 'ayah_audio/dukhain/001_001.mp3',
        'array': array([...]),  # Audio waveform as numpy array
        'sampling_rate': 16000
    },
    'text': 'بِسْمِ ٱللَّهِ ٱلرَّحْمَٰنِ ٱلرَّحِيمِ',
    'duration': 4.52,
    'reciter': 'dukhain'
}
```

### Data Fields

- **audio** (`Audio`): Audio file containing the recitation, automatically resampled to 16kHz
  - `path`: Original file path
  - `array`: Audio waveform as a 1-dimensional numpy array
  - `sampling_rate`: Sample rate (16000 Hz)
- **text** (`string`): Diacritized Arabic transcription of the verse
- **duration** (`float`): Duration of the audio clip in seconds
- **reciter** (`string`): Identifier for the reciter

### Data Splits

The dataset is randomly split into three subsets:

| Split | Number of Examples | Percentage |
|-------|-------------------|------------|
| Train | 17,460 | 70% |
| Validation | 3,741 | 15% |
| Test | 3,743 | 15% |

All splits maintain a balanced distribution of reciters and surahs.

### Reciters

This dataset includes recordings from four professional Quran reciters:

| Reciter ID | Reciter Name | Number of Ayahs | Source |
|------------|--------------|-----------------|--------|
| dukhain | Abdullah Matroud | 6,236 | EveryAyah.com |
| luhaidan | Abdullah Awad Al-Juhani | 6,236 | EveryAyah.com |
| salmi | Yasser Salama | 6,236 | EveryAyah.com |
| shahin | Mahmoud Khalil Al-Hussary | 6,236 | EveryAyah.com |

## Dataset Creation

### Source Data

#### Audio Sources
Audio recordings were obtained from [EveryAyah.com](https://everyayah.com), a platform providing high-quality Quran recitations by various reciters. Each reciter has a complete recording of the entire Quran.

#### Text Sources
The Arabic text with full diacritization (Tashkeel) was obtained from the [Quran JSON repository](https://github.com/semarketir/quranjson), which provides structured access to Quranic text.

### Data Collection and Processing

#### Collection Process
1. Downloaded full surah (chapter) recordings for each of the 4 reciters from EveryAyah.com
2. Obtained verse-level timing information in JSON format for precise segmentation
3. Downloaded diacritized Quranic text for all 114 surahs

#### Processing Pipeline
1. **Audio Segmentation**: Used ffmpeg to segment full surah recordings into individual verse-level audio files based on timing information
2. **Quality Control**: Verified that each audio segment meets minimum duration (0.5 seconds) and file size (1KB) thresholds
3. **Text Alignment**: Matched each audio segment with its corresponding diacritized Arabic text
4. **Audio Normalization**: Converted all audio files to 16kHz sampling rate and 64kbps MP3 format
5. **Dataset Splitting**: Randomly shuffled and split data into train/validation/test sets with fixed random seed for reproducibility

#### Who are the source data producers?
- Audio: Professional Quran reciters whose recordings are available on EveryAyah.com
- Text: The Quran JSON repository maintainers who structured the Quranic text
- Processing: Dataset compilation and processing by Rdyh

### Annotations

#### Annotation Process
The text annotations (Arabic transcriptions) are not manual annotations but rather the canonical Quranic text matched to each audio segment. The Quranic text is standardized and obtained from authoritative sources.

#### Who are the annotators?
Not applicable. The text is the canonical Quranic text, not manual transcriptions.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset is designed to support:
- Development of Arabic speech recognition systems specialized in Quranic recitation
- Building educational applications for Quran memorization and learning
- Research in classical Arabic phonetics and pronunciation
- Preservation of different recitation styles (Qira'at)

### Discussion of Biases

- **Reciter Diversity**: Limited to 4 male reciters; does not represent all recitation styles or genders
- **Audio Quality**: Recording quality may vary by reciter and original recording conditions
- **Dialect**: Contains only Classical Quranic Arabic, not modern dialects
- **Speaking Rate**: Recitation speed varies by reciter, which may affect model training

### Other Known Limitations

- Very short verses (under 1 second) may be challenging for ASR models
- Audio quality is dependent on original recordings from EveryAyah.com
- Limited to specific recitation styles; may not generalize to all Quranic recitations
- Does not include recitation rules (Tajweed) annotations beyond the diacritized text

## Additional Information

### Dataset Curators

This dataset was curated and compiled by Rdyh.

### Licensing Information

This dataset is released under the Creative Commons Attribution 4.0 International License (CC-BY-4.0).

### Citation Information

If you use this dataset in your research, please cite:

```bibtex
@dataset{everyayah_quran_2024,
  title={EveryAyah Quran Recitations Dataset},
  author={Rdyh},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/Rdyh/everyayah}}
}
```

### Contributions

Audio recordings courtesy of [EveryAyah.com](https://everyayah.com).
Quranic text from [Quran JSON repository](https://github.com/semarketir/quranjson).

## Usage Examples

### Loading the Dataset

```python
from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("Rdyh/everyayah")

# Load a specific split
train_dataset = load_dataset("Rdyh/everyayah", split="train")
test_dataset = load_dataset("Rdyh/everyayah", split="test")

# Access a sample
sample = dataset['train'][0]
print(f"Text: {sample['text']}")
print(f"Reciter: {sample['reciter']}")
print(f"Duration: {sample['duration']} seconds")
```

### Audio Processing

```python
import librosa
import numpy as np

# Access audio data
audio_array = sample['audio']['array']
sampling_rate = sample['audio']['sampling_rate']

# Process with librosa
mfcc = librosa.feature.mfcc(y=audio_array, sr=sampling_rate, n_mfcc=13)
```

### Training an ASR Model

```python
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

# Load pre-trained model and processor
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-large-xlsr-53-arabic")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-xlsr-53-arabic")

# Fine-tune on this dataset
# (training code here)
```

## Dataset Statistics

### Duration Distribution

- **Total Duration**: Approximately 174 hours
- **Average Duration per Ayah**: 25.1 seconds
- **Shortest Ayah**: 0.5 seconds
- **Longest Ayah**: 126 seconds (Ayat Al-Kursi and similar long verses)

### Coverage

Complete coverage of the Holy Quran:
- 114 Surahs (chapters)
- 6,236 unique verses
- 4 complete recitations
- 24,944 total audio-text pairs

### Text Characteristics

- Fully diacritized Classical Arabic
- Average verse length: ~65 characters
- Complete Tashkeel marks (Fatha, Damma, Kasra, Sukun, Shadda, Tanween)
- Includes special Quranic symbols (Hamza, Alif variations, etc.)