An Ontology learning model for Dysarthria speech using semantic-situational projection analysis

Abstract

Recent advances in Automated Speech recognition have succeeded in recognizing Dysarthria speech to a satisfactory measure. However, with the still ongoing research in these areas, the fundamental foundations of natural language communication, which is speech comprehension, still awaits to be met. This has remained a challenge due to the varied challenges posed by Dysarthria speakers such as irregularities in the speed, strength, pitch, breath control, breadth, composure, word omission, and accuracy of muscle motions required for speech production, which may be alien to an unfamiliar listener. This research acknowledged the duality of the Speech recognition task for Dysarthria speech as first having the speech being made intelligible and secondly, having the speech being put in the most likely context for purposes of meaning extraction. Whereas the existing approaches fall short of fully addressing the Speech Recognition problem in its duality, this research developed a model for speech comprehension for Speakers with Dysarthria that addresses the second half of the speech recognition task. The goal of this study therefore was to figure out how to create an ontology learning model that uses situational semantic projection to correctly comprehend and extract meaning from Dysarthric speech. This technique leveraged on word embeddings by representing words from Dysarthric speech as context vectors in a multidimensional space and analyzing the lexical co-occurrences, counterweights, and co-relation patterns in Dysarthric speech against those in natural language. This study was accomplished using a comparative experimental study approach that included quantitative data collecting via simulations of several algorithm versions and hyper-parameters. The suggested model was subsequently developed and tested through a series of experiments, employing both Dysarthric and non-Dysarthric speech data. Following triple extraction accuracy of 94%, the situational semantic projection model produced a 74% hit for a cosine closeness of 0.001 and a 55% correlation between the triple classes and the related emotional conversation actions. This ultimately resulted in a significant performance of above 0.5 (random guessing) for the ontology learning model given an accuracy of 77%.