Modelpublic

enver/tajweedsst

sourceHugging Faceupdated 8mo agoView on Hugging Face

1likes

Model Card

TajweedSST — Quranic Letter-Level Alignment & Tajweed Physics Engine

CTC Forced Alignment + Acoustic Physics Validation for Quranic Recitation

Overview

TajweedSST is a Python pipeline that produces letter-level timing data for Quranic recitation audio. It combines wav2vec2 CTC forced alignment with acoustic physics validation (Tajweed rules) to generate timing files consumed by MahQuranApp for real-time letter highlighting.

Pipeline Architecture

┌─────────────────────────────────────────────────────────────┐
│                    TajweedSST Pipeline                       │
│                                                             │
│  1. CTC Forced Alignment (wav2vec2)                         │
│     └─ Word-level timestamps from audio                     │
│                                                             │
│  2. Character Expansion                                     │
│     └─ Word timestamps → individual character timing        │
│                                                             │
│  3. Grapheme Matching                                       │
│     └─ Merge base + diacritics to match App.tsx rendering   │
│                                                             │
│  4. Tajweed Parsing                                         │
│     └─ Map letters to Tajweed rules (Qalqalah, Ghunnah..)  │
│                                                             │
│  5. Physics Validation                                      │
│     └─ RMS bounce, duration, formant analysis               │
│                                                             │
│  6. Export to MahQuranApp format                             │
│     └─ JSON with idx, char, ayah, start(ms), end, wordIdx   │
└─────────────────────────────────────────────────────────────┘

Quick Start

Prerequisites

bash

cd /path/to/tajweedsst
python3 -m venv venv
source venv/bin/activate
pip install torch torchaudio ctc-forced-aligner librosa

Single Surah

bash

# Align Surah 90 (Al-Balad) for Abdul Basit
python ctc_align_91.py  # Template script

Batch All Surahs

bash

# Process all 114 surahs for Abdul Basit
python batch_align_all.py

Output Format

Each letter_timing_XX.json contains an array of timing entries:

json

{
  "idx": 0,
  "char": "لَ",
  "ayah": 1,
  "start": 3360,
  "end": 3410,
  "duration": 50,
  "wordIdx": 0,
  "weight": 1.0
}

Fields

Field	Type	Description
`idx`	int	Sequential letter index
`char`	string	Arabic grapheme (base + diacritics)
`ayah`	int	Verse number (1-indexed)
`start`	int	Start time in milliseconds
`end`	int	End time in milliseconds
`duration`	int	Duration in milliseconds
`wordIdx`	int	Word index within the surah
`weight`	float	Confidence weight

Critical: Grapheme Matching

The timing data must match the grapheme count produced by MahQuranApp's splitIntoGraphemes() function. This function combines base Arabic letters with their following diacritics:

App.tsx Diacritics Set:

ً ٌ ٍ َ ُ ِ ّ ْ ٰ ۖ ۗ ۘ ۙ ۚ ۛ ۜ ٔ ٓ ـ

Plus Unicode ranges: 0x064B–0x0652 and 0x0610–0x061A

Example: The word لَآ splits into 2 graphemes: ['لَ', 'آ']

If the timing count doesn't match the grapheme count, highlighting will drift!

Physics Validation

TajweedSST validates timing against acoustic physics:

Rule	Check	Method
Qalqalah	RMS dip + spike	Envelope analysis
Ghunnah	Nasal duration	Duration measurement
Madd	Extended vowel	Duration ratio
Tafkheem	Heavy articulation	Formant F2 analysis

Project Structure

tajweedsst/
├── src/
│   ├── tajweed_parser.py     # Tajweed rule detection
│   ├── physics_validator.py  # Acoustic validation
│   └── duration_model.py     # Duration calibration
├── tests/                    # 34 unit/integration tests
├── ctc_align_90.py          # Single surah alignment
├── ctc_align_91.py          # Template with physics
├── batch_align_all.py       # Batch all surahs
└── README.md

Reciter Support

Currently supported:

—Abdul Basit (114 surahs)

License

MIT