somezay/everyayah-masjid-augmented

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.

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:

{
  "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 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.