May 2026 · Posted · Final year project
DyslexAI — screening from handwriting and speech
Screening a child for dyslexia usually means waiting months for a specialist. This reads their handwriting and listens to them read, and says whether it's worth making the appointment.
As posted
Detect · decide · actReversed letters, erratic spacing, halting reading.
A risk band from independent handwriting and speech signals.
Issues a report, and says when to see a specialist.
The problem
Dyslexia is most treatable when it's caught early, but the indicators — reversed letters, erratic spacing, halting reading — are the kind of thing a parent or teacher notices without knowing whether it means anything. The assessment that would tell them sits behind a specialist waiting list.
The approach
Two modalities scored independently, then combined. A handwriting image is deskewed and contrast-normalised, then run through six detectors — letter reversals, spacing, size consistency, baseline deviation, stroke width variance and slant — with an EfficientNet-B0 CNN, trained on the Kaggle dyslexia handwriting set, classifying individual letters as normal, reversed or corrected. A recording of the Rainbow Passage yields five more signals: reading speed, word error rate, pauses, repetitions and phoneme errors, with Groq's Whisper Large v3 as a fallback when browser transcription drops out, guarded against hallucinated output. The signals become one risk band and a downloadable report. It is deliberately a screening tool, not a diagnosis — a positive result is an argument for seeing a specialist, and the report says so.