Noor

AI that learns how dyslexia affects the way you read the Qur’an — and adapts the page around you.

Dyslexia can make Qur’an reading especially difficult because Arabic requires readers to process connected letters, changing letter forms, dots, diacritics, sound patterns, and letter position at the same time.

But dyslexia is not one single problem. Different readers may struggle with visual tracking, phonological decoding, grapheme confusion, crowding, or letter-position errors.

Most digital tools only see the final mistake.

Noor tries to understand the pattern behind the mistake.

What Noor Does

Noor combines:

• webcam gaze tracking
• Qur’an-specific phoneme recognition
• Qur’anic word and grapheme structure
• timing and hesitation data

These signals form a Multimodal Reading Trace.

eyes + voice + Qur’an text + timing
                ↓
        reading trace
                ↓
       pattern detection
                ↓
      targeted intervention
Cognitive Pattern Matching

Noor uses deterministic rules to match observable reading behavior to dyslexia-related reading patterns.

Pattern What Noor detects   Intervention
Visual tracking instability Gaze moves ahead/behind the spoken word, repeated regressions   Progressive reveal + faded surrounding text
Visual crowding sensitivity Errors increase when nearby text is dense   Reduce visual density + isolate phrase
Graphemic confusability Repeated confusion between similar Arabic letters   Highlight dots and distinguishing features
Phonological decoding difficulty    Long fixation + hesitation + incorrect phoneme  Grapheme-to-phoneme synchronization
Diacritic decoding difficulty   Consonants correct, harakah-related sounds repeatedly wrong Emphasize diacritics + synchronized sound
Letter-position errors  Correct sounds appear in the wrong order    Sequential grapheme reveal

Noor does not diagnose dyslexia. It detects observable reading patterns and assigns support accordingly.

Personalized Learner Profile

Each session builds a profile of strengths and support areas.

Reading dimension   Example
Phoneme accuracy    88%
Harakah decoding    82%
Word completion 79%
Visual tracking 32%
Graphemic discrimination    41%

The goal is simple:

What is this reader good at, where do they struggle, and what actually helps them improve?

How We Built It
Gaze Estimation

A webcam gaze model estimates where the reader is looking and maps that position to Qur’an words.

This gives us:

fixation duration
regressions
gaze progression
gaze–voice divergence
Qur’an Phoneme Recognition

A specialized Qur’an phoneme model captures what sounds the reader actually produced.

We compare those phonemes against the expected Qur’anic sequence to detect:

substitutions
omissions
repetitions
pauses
letter-order errors
Deterministic Pattern Engine

The system applies explicit rules.

gaze–voice divergence
+
multiple regressions
        ↓
visual tracking instability
        ↓
progressive reveal

This makes Noor explainable and predictable instead of relying on an LLM to guess the learner’s difficulty.

Why Noor Is Different

Most learning software adapts to whether you were right or wrong.

Noor adapts to how you struggled.

Two readers can make the same mistake for different reasons, so Noor can assign different interventions to each.

same mistake
→ different reading trace
→ different cognitive pattern
→ different support

Every reader struggles differently. Their learning tools should too.

Built With

• ml
• python