somezay/everyayah-masjid-augmented

---
license: cc-by-4.0
language:
- ar
pretty_name: EveryAyah Mosque-Environment Augmented
size_categories:
- 10K<n<100K
task_categories:
- automatic-speech-recognition
tags:
- quran
- arabic
- mosque
- audio-augmentation
- whisper
- fine-tuning
- asr
---

# EveryAyah — Mosque-Environment Augmented Dataset

Quran recitation audio with realistic mosque acoustic augmentation applied, designed for fine-tuning ASR (Automatic Speech Recognition) models that need to work in real mosque/masjid environments.

## Why This Dataset Exists

Standard Quran recitation datasets (like EveryAyah) are recorded in studio conditions. ASR models trained on clean audio perform poorly in real mosques due to:

- **Heavy low-pass filtering** — mosque rooms act as natural LPFs (Low-Pass Filters), with 97% of energy below 500 Hz
- **Reverberation** — large open prayer halls create RT60 times of 1-2 seconds
- **Low-frequency resonance** — bass buildup from room modes around 80-200 Hz
- **Ambient noise** — air conditioning hum, shuffling, breathing

This dataset applies empirically-matched augmentation based on spectral analysis of real mosque recordings (spectral rolloff measured at 478 Hz, -20 dB/octave above 500 Hz).

## Dataset Structure

```
masjid_medium/{reciter}/{SSS_AAA}.mp3 — moderate mosque conditions
masjid_heavy/{reciter}/{SSS_AAA}.mp3 — harsh mosque conditions
manifest_upload.jsonl — metadata (path, text, surah, ayah, reciter, augmentation)
```

### File Naming

- `SSS` = 3-digit surah number (001-114)
- `AAA` = 3-digit ayah number

### Augmentation Presets

| Preset | LPF Cutoff | Reverb RT60 | Wet Mix | Bass Boost | Gaussian SNR | Hum SNR |
|--------|-----------|-------------|---------|------------|-------------|---------|
| **masjid_medium** | 600 Hz | 1.2s | 0.28 | +4 dB @ 120 Hz | 28 dB | 28 dB |
| **masjid_heavy** | 500 Hz | 1.7s | 0.40 | +5 dB @ 120 Hz | 25 dB | 25 dB |

The augmentation chain applies (in order):
1. Reverb (synthetic impulse response)
2. Low-frequency resonance boost
3. Gaussian noise
4. Low-frequency hum (50/60 Hz harmonics)
5. Low-pass filter (simulating room acoustics)

### Reciters (from EveryAyah.com)

| Reciter | Clips per augmentation |
|---------|----------------------|
| Alafasy_128kbps | 6,236 |
| Husary_128kbps | 6,235 |
| Abdul_Basit_Murattal_192kbps | 6,234 |
| MaherAlMuaiqly128kbps | 6,236 |
| Minshawy_Murattal_128kbps | 6,228 |

**Total: ~62,338 augmented clips** across 2 augmentation levels.

## Clean Audio

Clean (unaugmented) audio is not included — download directly from [EveryAyah.com](https://everyayah.com/).

A lighter augmentation level (`masjid_light`: LPF@800Hz, SNR 32dB) was also generated but excluded from this upload as it's close enough to clean audio to be less useful for training.

## Manifest Format

Each line in `manifest_upload.jsonl` is a JSON object:

```json
{
"path": "masjid_heavy/Alafasy_128kbps/001_001.mp3",
"text": "بِسْمِ اللَّهِ الرَّحْمَنِ الرَّحِيمِ",
"surah": 1,
"ayah": 1,
"reciter": "Alafasy_128kbps",
"augmentation": "masjid_heavy"
}
```

## Intended Use

- Fine-tuning Whisper (or other ASR models) for mosque environments
- Training noise-robust Quran recitation recognizers
- Benchmarking ASR robustness to room acoustics

## How It Was Made

1. Downloaded all ayah-level MP3s from [EveryAyah.com](https://everyayah.com/) for 5 reciters
2. Decoded to 16 kHz mono WAV
3. Applied augmentation chain calibrated against real mosque recordings
4. Re-encoded to 128 kbps MP3

Spectral calibration was done by comparing synthetic augmentation output against real Tarawih prayer recordings captured on a Galaxy A33 phone placed on the mosque floor.

## Citation

If you use this dataset, please credit EveryAyah.com as the original audio source.

## License

Audio content: recordings from EveryAyah.com. Augmentation and metadata: CC-BY-4.0.