Speech-to-Text Based Dyslexia Detection Using Deep Learning and Visual Pattern Analysis

Dyslexia is a neurodevelopmental condition affecting 5-10% globally, marked by persistent difficulties in reading accuracy, spelling, and word decoding despite normal intelligence.Early detection is crucial for timely support, but traditional screenings are labor-intensive, subjective, and often delayed until academic struggles appear.This study proposes a framework that combines speech-to-text processing with visual analysis of transcribed speech.It assesses both the quality of automated transcription and patterns linked to dyslexia.The approach uses a multimodal deep learning system: a CNN-LSTM hybrid converts children's reading aloud audio into text, while a transformer-based model analyzes temporal and visual cues of hesitations, mispronunciations, and self-corrections.Training and evaluation were performed on 847 audio samples from children aged 6-12 years.Experimental results show 94.3% classification accuracy, 92.7% sensitivity, and 95.1% specificity, indicating strong diagnostic performance.Beyond prediction, the framework produces interpretable outputs that identify specific phonetic and timing difficulties, helping educators and speech-language pathologists tailor interventions.Overall, speech-to-text analytics with visual behavior modeling offers a scalable, non-invasive, practical tool for early dyslexia screening, complementing conventional methods.The dataset spans grades 1-5, covering varied reading abilities, and the models prioritize interpretability alongside accuracy in practice.