Improving Intelligibility of Dysarthric Speech using Landmark Detection and Modifications on Distorted Stops

Abstract

Dysarthria is a motor speech disability in which people have difficulties with articulators
used in the development of speech. The lack of control over the articulators affects distorted
speech whose intelligibility, relative to regular speakers, is very low. Because the distortion is
dependent on the type of articulator and severity of the disorder, the issues are usually unique
to the person. The issue of devoicing in voiced stops is discussed, as well as phonemic
distortions in the database of dysarthric speech for universal access research. Automatically
detecting landmarks determines the segments that need to be modified. Before the burst
onset, a voice bar of low frequency is added to correct devoicing in voiced stops. Frication
due to incomplete closure during the production of stops was identified and frication in the
closure is removed. As dysarthric speech is slower than normal speech, these modifications
are accompanied by time scaling. After these modifications, a subjective intelligibility
assessment is done. The intelligibility assessment by the eight participants showed that both
modifications contribute to improving the intelligibility of mispronounced words.
Transcriptions for devoiced stops improved from 1.56% to 95.31%, and for incomplete
closure from 0% to 80%, after the modification.