AhmedTamertechno1/Phone_Timings_Database

📖 TajweedAI: Quranic Phoneme Timing Benchmark (Phases 1, 2 & 3)

📌 Project Overview

TajweedAI evaluates Quranic recitation accuracy by analyzing both pronunciation (phoneme classification) and timing (rule duration evaluation).

This benchmark provides empirical, tempo-normalized duration boundaries for all 70 Quranic phonemes derived from forced alignments (MFA trained on Quranic audio) across 7 master reference reciters:

1. Sheikh Mahmoud Khalil Al-Husary (Gold Standard Reference)
2. Dr. Ayman Sowaid (Pedagogical Reference)
3. Sheikh Abdullah Basfar
4. Sheikh Abu Bakr Ash-Shaatree
5. Sheikh Mishary Rashid Alafasy
6. Sheikh Saad Al-Ghamadi
7. Sheikh Maher Al-Muaiqly

---

🛠️ Methodology & Phases

Phase 1: Log-IQR Outlier Filtering

• Quranic phoneme durations are right-skewed.
• Transformed durations into log space: $y = \ln(t)$.
• Applied Log-IQR filtering: $[Q_1 - 1.5 \times \text{IQR}, Q_3 + 1.5 \times \text{IQR}]$ to eliminate frame glitches while preserving natural linguistic variance.

Phase 2: Tempo Normalization Against Master Reference

• Standardized varying recitation speeds against Sheikh Al-Husary (and Dr. Ayman Sowaid for rare geminated tokens): $$S_{R,p} = \frac{\text{Median}(\text{Reference}, p)}{\text{Median}(R, p)}$$
• Scaled all tokens: $t_{\text{norm}} = t_{\text{filtered}} \times S_{R,p}$.

Phase 3: Unified Token Pooling & Boundary Extraction (Option A)

• Pooled 289,591 tempo-normalized tokens across all 7 Imams into unified distributions for all 70 phonemes.
• Lower Bound Cutoff ($Min_{\text{valid}}$): $P_1$ (1st percentile) to remove 10–20 ms alignment boundary noise.
• Target Center: $P_{50}$ (Median, aligned to reference).
• Core Ideal Tier: $[Q_1, Q_3]$ (25th to 75th percentiles).
• Custom Madd Vowel Rules (Option A):
  • **a: (Alif Madd):** Full empirical maximum preserved ($Max = 2,944.8\text{ ms} = 2.945\text{ s}$).
  • **i: (Yaa Madd):** Highest integer percentile $< 4.0\text{ s}$ ($P_{90} = 3,726.6\text{ ms} = 3.727\text{ s}$).
  • **u: (Waaw Madd):** Highest integer percentile $< 4.0\text{ s}$ ($P_{97} = 3,828.0\text{ ms} = 3.828\text{ s}$).
  • Other 67 Phonemes: Standard $P_{99}$ upper bound.

---

📊 Tajweed Grading Tier Definitions

Tier	Range	Tajweed Meaning
🟢 Green Tier (Master / Ideal Core)	$[Q_1, Q_3]$	Ideal central target timing window
🟡 Yellow Tier (Permissible Tolerance)	$[Min_{\text{valid}}, Max_{\text{valid}}]$	Valid allowable recitation timing
🔴 Red Tier (Tajweed Violation)	$< Min_{\text{valid}}$ or $> Max_{\text{valid}}$	Under-elongated (dropped letter) or excessive elongation (Tamṭīṭ)

---

📂 Repository Contents

• phase3_unified_phoneme_timings_optionA.csv: Master final lookup table for all 70 phonemes.
• all_reciters_phase3_pooled_tokens.csv: Complete pooled token dataset ($N = 289,591$).
• phase2_scaling_factors.csv: Reciter scaling factor matrix.
• timing_lookuptable.md: Human-readable Markdown reference table.
• Phase3_Unified_Plots/: 70 unified distribution histograms with annotated threshold lines.
• Phase3_Boxplots_Vowels/: Per-Imam and pooled boxplots for Madds (a:, i:, u:).
• Phase2_Comparative_Plots/: 70 multi-reciter comparative before/after normalization plots.
• Scripts/: Reproducible Python scripts (run_phase1.py, run_phase2.py, run_phase3.py).