DySARNet: a lightweight self-attention deep learning model for diagnosing dysarthria from speech recordings

Dysarthria, a neurological condition impairing speech, has garnered significant research interest in Deep Learning (DL). Curent studies underscore DL’s potential in automating dysarthria diagnosis and severity classification, utilizing well-known benchmark datasets like TORGO and UA-Speech. However, existing DL for dysarthria diagnosis, like Deep Neural Networks, are often computationally expensive, complex, and lack context awareness. This paper addresses these issues by proposing a Densely Squeezed-and-excited Attention-gated Residual convolutional neural Network (DySARNet) designed for dysarthria diagnosis and severity estimation. DySARNet aims to provide a cost-effective solution compatible with various devices, including mobile phones. DySARNet made this possible by integrating a Lean Separable Dense Block (LSDB) that enhances feature reuse while preventing a massive parameter increase through separable depthwise convolutions (SDWConv). For improved context understanding and awareness, DySARNet employs a Squeeze-and-Excited Lightweight Residual Attention-gated (SELRA) mechanism, a tailored self-attention mechanism that offers robust attention at a low cost by squeezing additional parameters through SDWConv and depthwise convolutions. DySARNet achieves 99.03% and 98.77% accuracy in diagnosing dysarthric and non-dysarthric speech recordings and 99.20% and 96.09% in classifying dysarthria severity using the TORGO and UA-Speech datasets. With only 33 K parameters and 0.04 G-FLOPs, DySARNet surpasses most state-of-the-art models in cost-effectiveness and overall performance. In conclusion, DySARNet, with its SELRA and LSDB components, reveals significant promise in enhancing dysarthria diagnosis by automatically recognizing the presence of dysarthria and classifying its severity more efficiently.