Optimizing Vocabulary Modeling for Dysarthric Speech Recognition

Abstract

Imperfection in articulation of dysarthric speech results in the deterioration on the performance of speech recognition. In this paper, the effect of the articulating class of phonemes in the dysarthric speech recognition results is analyzed using generalized linear mixed models (GLMMs). The model with the features categorized according to the manner of articulation and the place of tongue is selected as the best one by the analysis. Recognition accuracy score for each word is predicted based on its pronunciation and the GLMM. The vocabulary optimized by selecting words with the maximum score shows a 16.4 % relative error reduction in dysarthric speech recognition.