A novel class-attention transformer-driven feature fusion technique-based speech disorder classification

Speech disorders present a diagnostic challenge due to complex acoustic characteristics in speech signals. The study proposes a novel system for detecting pathological speech disorders that processes raw audio signals and distinguishes between disordered and healthy speech. The method combines one-dimensional convolutional neural networks with a transformer and self-attention mechanism. The model achieves 97.50% accuracy on two benchmark datasets (SVD and PD) and 95.80% accuracy on the VOICED dataset, while maintaining a lightweight design with 5.2 million parameters. The system enables visualization and interpretation of model decisions, helping clinicians and speech-language pathologists identify diagnostically significant time intervals. The results were verified through statistical analysis and tested on multiple independent datasets.