Clinically Interpretable Hybrid Deep Neural Architecture for Multi-Domain Acoustic Feature Learning in Dysarthric Speech Analysis

Dysarthria is a speech impairment that significantly disrupts communication and can emerge as an early sign of neurological diseases such as ALS and Parkinson's disease. This paper introduces an interpretable deep learning framework that synergistically integrates Convolutional Neural Networks (CNNs) with Long Short-Term Memory (LSTM) networks to automatically detect dysarthria using Mel-Frequency Cepstral Coefficients (MFCCs) and temporal acoustic features. The frame-work effectively captures both spectral and temporal aspects of speech. It achieves an accuracy of 0.97, F1-score of 0.97 and an AUC of 0.99. We use Explainable AI (XAI) techniques like LIME and Grad-CAM, which are employed to visualize key acoustic areas responsible for classification. Model evaluation and feature contribution studies demonstrate the model's dependability and efficiency. The implemented architecture shows promise for use in clinical applications and digital health. It offers a computational solution for timely and unbiased detection of dysarthric speech in practical medical scenarios.