SARNet: Speaker-Attentive ResNet for Quantification of Dysarthria Severity

This is the authors' abstract. We don't add key points for this paper.

Analyzing differences in audio data of individuals with articulation disorders from the perspective of human speech and employing objective methods for automated dysarthria evaluation can significantly aid doctors in early patient screening and diagnosis. This proactive approach enables timely intervention and treatment during the initial stages of the condition. This paper presents a speech recognition network(named Speaker-Attentive ResNet (SARNet)) that aggregates and propagates features from different hierarchical levels using an attention-based statistical pooling module, built upon the ResNet architecture. This network aims to extract subtle features specific to individual speakers. The proposed method is evaluated using the TORGO dysarthric speech database, employing eight different acoustic features as well as features aggregated from these eight. Comparative experiments with two other models, Time-Delay Neural Network (TDNN) and Panns-CNN10(Large-Scale Pretrained Audio Neural Networks), demonstrate that due to the superior ability of MFCC(Mel-scaleFrequency Cepstral Coefficients) to emulate human auditory features, the proposed approach achieves a classification accuracy of $98 %-99 %$ in measuring speech intelligibility in healthy people, patients, and the severity of patient’s articulation disorders.