DSSCNet: A Transfer Learning Framework for Cross-Corpus Dysarthric Speech Severity Classification

Researchers introduce DSSCNet, a deep learning framework using transfer learning to improve dysarthric speech severity classification across different datasets. The model achieves 75.80% accuracy on TORGO and 68.25% on UA-Speech corpora, demonstrating significant improvements in speaker-independent assessment and cross-corpus generalization for assistive speech technologies.

DSSCNet addresses a critical gap in medical AI by tackling dysarthric speech classification, a problem where traditional machine learning struggles due to speaker variability, imbalanced datasets, and limited training data. Dysarthria affects millions globally, making automated severity assessment valuable for clinical diagnosis and monitoring. The research demonstrates that transfer learning—pre-training on one speech corpus and fine-tuning on another—enables models to generalize better across populations and recording conditions, a challenge that has hindered deployment of speech technology in healthcare settings.

The field of speech pathology has historically relied on subjective clinician evaluations, creating bottlenecks in diagnosis and treatment planning. DSSCNet's multi-corpus learning approach mirrors successful transfer learning applications in computer vision and natural language processing, suggesting that healthcare AI can benefit from cross-dataset knowledge transfer. The reported accuracy improvements over state-of-the-art baselines indicate the framework's practical viability.

For healthcare developers and assistive technology companies, DSSCNet reduces implementation barriers by demonstrating that models trained on limited dysarthric speech data can still achieve robust performance through intelligent transfer learning. This enables smaller organizations to develop diagnostic tools without requiring massive proprietary datasets. The framework also has implications for under-resourced healthcare settings where specialist speech pathologists are scarce.

Future adoption depends on clinical validation studies, regulatory approval, and integration into existing clinical workflows. Researchers should investigate how DSSCNet performs across different dysarthria etiologies (stroke, cerebral palsy, Parkinson's) and whether real-time deployment on edge devices remains feasible. Commercial viability hinges on demonstrating clinical equivalence to human experts and achieving regulatory clearance.

• →Transfer learning enables dysarthric speech models to generalize across different corpora and speaker populations effectively
• →DSSCNet achieves 75.80% accuracy on TORGO dataset, outperforming previous state-of-the-art classification methods
• →Multi-corpus learning reduces reliance on large annotated datasets, lowering barriers for healthcare AI development
• →The framework addresses speaker-independent classification, making automated dysarthria severity assessment clinically practical
• →Cross-dataset knowledge transfer demonstrates that healthcare AI can adopt techniques proven in other AI domains