Attention-based Deep Feature Fusion for Automated Dysarthria SeverityClassification: A Speech-based Computational Functional Marker Relevantto Neurovascular and Neurodegenerative Conditions

introduction: Dysarthria is a motor speech disorder that frequently arises as a clinical manifestation of neurovascular and neurodegenerative diseases, including stroke, traumatic brain injury, Parkinson’s disease, and amyotrophic lateral sclerosis. These conditions disrupt neuronal and vascular pathways involved in motor speech control, leading to impaired articulation, delayed speech, and reduced intelligibility. As dysarthric speech is delayed and imprecise, the degree of the impairment determines speech intelligibility. This establishes a communication barrier between dysarthric patients and the public. materials and methods: To this aim, Deep Learning (DL) methods are proposed for automated severity classification of dysarthria. This work employs different methodologies to identify the optimal setup for classifying severity in dysarthria. The first two methods utilize baseline Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) with cepstral features as input. The third method experiments with 6 pre-trained networks for automated severity classification. The fourth experiment utilizes deep spectrum features extracted from dysarthric Mel-spectrograms of pre-trained networks followed by Support Vector Machine (SVM) classification. The final experiment employs attention-based fusion of deep spectrum features obtained from top -2 pre-trained networks and classified by SVM and its variants. results: An accuracy of 97.90 % and 95.31 % is obtained in Torgo and UA databases, respectively using attention based deep feature fusion. discussion: Various experimentations reveal that optimal choice of DL framework using pretrained networks as feature extractors in conjunction with feature engineering can improve the performance of severity classification of Dysarthria. conclusion: . The severity of dysarthria reflects the extent of underlying neuronal damage, making it a valuable biomarker for monitoring disease progression and guiding neurorehabilitation strategies.