Towards Dys-XAI: Influence-Based Explanations for Dysarthria Severity Assessment

The article presents an influence-based explainability framework for assessing dysarthria severity, addressing the limitations of existing speech explainability methods. By utilizing gradient-based influence approximations, the framework computes per-utterance influence scores to identify relevant training samples for each prediction, enhancing interpretability for clinicians. This method improves the clinical adoption of deep learning models by providing auditable explanations that connect model decisions to perceptible reference cases, which is crucial for therapy planning and monitoring in dysarthria.