Landmark based modification to correct distortions in dysarthric speech

Dysarthria is a motor speech disorder in which the person lacks the control over articulators used for speech production. This results in distorted speech, whose intelligibility is poor compared to the normal speakers. The distortion depends on the type of articulator and severity of the disease, because of which the problems tend to be person specific. Phonemic distortions in the dysarthric speech database for universal access research are analyzed in this work and the problem of devoicing in voiced stops is addressed. In this work, segments for modification are identified by automatically detecting the landmarks. Devoicing in voiced stops is rectified by introducing voice bar of low frequency before the burst onset. Intelligibility assessment is carried out by subjective test where, listeners assessed the intelligibility. Large improvement in intelligibility is found after modification.