AI-Assisted Oral Reading Screening — Flagging Reading Difficulty in Greek Primary Pupils

Researchers at the University of Thessaly (Departments of Informatics & Telecommunications and Special Education) developed a deep-learning screening tool that analyses 7-second audio clips of Greek primary-school pupils reading aloud to flag potential reading difficulty.

The tool is designed as a low-burden screening aid for teachers — not a diagnostic instrument — to help identify pupils who may need further evaluation for reading difficulties or dyslexia.

Teachers record pupils reading aloud; the audio is segmented into 7-second clips, converted to spectrograms, and classified by a YOLO-based convolutional neural network, producing an aggregate "difficulty percentage" score for each pupil.

In a validation sample of 77 pupils (grades 3-6, mean age 11.6, including 6 with dyslexia diagnoses), the tool achieved about 84% overall accuracy, 0.85 balanced accuracy, and 0.89 macro F1, with a 0.74 correlation to expert judgment, using an 80/20 participant-level data split.

The authors state the tool needs further validation — including classroom-noise robustness testing, larger and more diverse samples, and multilingual testing — before wider implementation, and stress it should function only as a screening aid rather than a diagnostic tool.