Dysarthria Severity detection Using Recurrent and Convolutional Neural Networks

The diagnosis and monitoring of dysarthria, a speech disorder caused by neuro-motor problems that affect articulation, depend on a precise evaluation of its severity. When creating automated systems to identify and categorize dysarthric speech, accurate severity classification is essential. Using neural network models, specifically recurrent neural networks (RNN) and convolutional neural networks (CNN), this paper offers a thorough investigation of how to distinguish dysarthric voices among a collection of normal voice samples and categorize the severity of dysarthria. Among the features used in the study are voice quality, prosodic parameters, formants, Mel frequency cepstral coefficients (MFCC), and spectrograms. Comparing the ability of convolutional networks and reccurents to identify abnormalities in normal data, as well as the hybrid model that combines convolutional and reccurent neural networks (CRNN), is our goal. The Nemours corpus database is used to assess these neural network models' performances. Notably, 99.8% is the highest classification accuracy attained with this corpus