Approaches for Controlled / Dysarthric Speech Classification Based on Sinusoidal Model

Dysarthria is a speech disorder which is caused by the poor co-ordination of muscles which are involved in speech production mechanism. Diagnostic methods can be supported by various machine learning and deep learning models, depending on the nature and variability of dysarthric speech, including its severity. This paper focuses on efficient dysarthria classification methods based on sinusoidal model of speech. Dysarthric speech has disturbed harmonic structure compared to normal speech and this is expected to get well depicted in the frame to frame discontinuity of parameters of sinusoidal model. Three different models are suggested in this paper which can significantly contribute for more accurate classification of controlled/dysarthric speech. With amplitudes and frequencies of sinusoids as feature vectors for baseline classifiers and Deep Neural Networks (DNN), the proposed models provide an accuracy in the range of 70-85%. The paper also illustrates variations in the utterance level SNR of TORGO and UA-Speech databases with the help of intensity threshold detection on an experimental basis. The sinusoidal model’s effectiveness in speech analysis is also evaluated by adjusting the number of peaks that account for the distinct noise profiles of the UA-Speech and TORGO datasets.