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

Introduction: Speech disorders (SD) present significant diagnostic challenges due to the complex and indistinct acoustic characteristics embedded within speech signals. Problem statement: The standalone convolutional neural networks (CNNs) and vision transformers (ViT) struggle to capture SD temporal and spectral dynamics. Objectives: To address these limitations, this study proposes an end-to-end binary pathological SD detection framework that processes raw audio waveforms to differentiate disordered speech from healthy speech while improving feature representation, interpretability, and generalization. Methods: A hybrid feature extraction integrating one-dimensional CNNs and ViT's self-attention mechanism is developed to extract crucial SD features. An adaptive fusion and a Class-attention transformer (CaiT)-based feature refinement is introduced to transform the extracted features into a classification-optimized global representation. Through the integration of gradient-weighted class activation mapping (Grad-CAM) and attention-based visualization with the model architecture, the temporal-localization of disorder-relevant acoustic patterns is enabled. Two benchmark datasets are utilized for model's performance evaluation, benchmarking against state-of-the-art CNNs and ViT architectures. The model is trained and internally validated on the SVD dataset using subject-level splitting, while the PD and VOICED datasets are used exclusively for external validation to assess cross-dataset generalization across neurological and heterogeneous pathological voice conditions. Results: The proposed model achieves 97.50% accuracy on both SVD and PD datasets and 95.80% accuracy on VOICED, while maintaining a lightweight design with 5.2 million parameters. Statistical analysis further confirms the results' reliability and significance. Conclusion: The integration of adaptive fusion and transformer-based refinement significantly enhances SD detection performance while ensuring interpretability. The suggested framework offers a sound and proof-of-concept for diagnosing SD, allowing clinicians and speech-language pathologists to make reliable predictions and pay attention to diagnostically significant time intervals.