Speak-up: web-based speech disorder detection system with recommended speech therapy for grade four and five elementary students using Convolutional Neural Network

Detecting speech disorders in children is challenging for existing tools as they rely on the resource-intensive assessments performed by Speech Language Pathologists (SLP). Similarly, speech therapy is often inaccessible and expensive, especially in rural communities. To address these concerns, the researchers developed the following: an automated acoustic feature extraction for both live and recorded audio input. Second, a web-based system that can classify normal and disordered speech namely articulation disorder, phonological disorder, and stuttering. Third, gamified speech therapy exercises that administer early interventions for children. Using Convolutional Neural Network (CNN), the model achieved 88% accuracy in identifying normal and disordered speech. Normal and stuttering achieved F1 scores of 93% and 91% respectively, while articulation and phonological disorders achieved F1 scores of 84%. Following the model’s integration into the web-based system, it was evaluated on 34 children with identified speech disorders. The model correctly identified stuttering with 81% accuracy, phonological and articulation disorders with 71% and 73% accuracy, respectively, for an overall accuracy of 75%. Furthermore, the researchers used the ISO 25010:2011 software quality standard to evaluate the system. The accumulated results were satisfactory in terms of functional suitability, performance efficiency, usability, and reliability, with mean scores of 3.14, 3.19, 3.10, and 3.14, respectively.