Linear Prediction Residual Error Based Assessment of Dysarthric Speech

Abstract

Dysarthria is a neuromotor impairment of speech that affects one or more subsystems like laryngeal, velopharyngeal and articulatory. The acoustic characteristics of dysarthric speakers deviate from that of the normal speakers. For the assessment of velopharyngeal dysfunction (hypernasality) inclusion of additional formants and deletion of actual formants are considered as acoustic cues. In this work, a novel approach is proposed for the assessment of dysarthric speech, using linear prediction residual error. Speech signal of dysarthric speakers is passed through inverse filter, designed using the system parameters estimated using LP analysis of a normal speaker, which is the LP residual error, is expected to contain information about the additional resonance frequency introduced due to hypernasality. On the other hand, the reverse process is expected to give the deletion of formants. In order to perform this analysis, an acoustically closer (in terms of formant frequency) normal speaker is to be found. For this comparison, the distance metrics like accumulated residual energy, Itakura-Saito distance and COSH distance measures are used. The LP-based magnitude spectra of the residual error, derived as mentioned above, contain information about the deletion and addition of resonant frequencies. Acoustic analysis is performed on 18 hypernasal speaker‟s speech data collected from AIISH, Mysore and 15 normal speakers‟ speech data collected in the laboratory environment and 10 dysarthric speakers‟ data from Nemours database of dysarthric speech. It is observed that out of 10 dysarthric speakers of Nemours database, four speakers are found to be hypernasal.