Dysarthria Diagnosis and Dysarthric Speaker Identification Using Raw Speech Model

Dysarthria is a medical condition that causes difficulty in producing coherent speech due to muscle paralysis or weakness. This article presents a unique approach to identifying dysarthric speakers using a deep learning model that works directly with unprocessed speech waveforms. By eliminating the need for feature extractions, the model's resistance to noise and voice variability is increased. The proposed approach utilizes a SincNet layer model with multiple initializations including Mel, Erb, and Bark scales for dysarthria detection (DD) and dysarthric speaker identification (DSI). Bark scaling, aligning better with human auditory perception and capturing distinctive acoustic features, notably outperforms other initialization methods. When Bark scaling was employed, the study's results demonstrated outstanding accuracy rates of 97.0% for DD and 88.0% for DSI. The results demonstrated exceptional performance, that surpassing existing literature benchmarks.