Significance of Feature Selection for Acoustic Modeling in Dysarthric Speech Recognition

Abstract

In this paper, a comparative study of various feature extraction methods is carried out on dysarthric speech. Dysarthric speech is difficult to recognize and thus pose challenges that normal speech does not. Since various features can be used to model phonemes in hidden Markov model (HMM) based recognition system, which feature is suitable for the task specified is a topic to be addressed.Dysarthric speech becomes unintelligible due to the improper coordination of articulators. In this paper, recognition results are compared using mel-frequency cepstral coefficients (MFCC), perceptual linear prediction (PLP), filter bank and reflection coefficients feature sets. The performance is analyzed using TORGO database. Phonemes are grouped for the analysis. Our study shows that MFCC and PLP gave better results than filter bank and reflection coefficients for dysarthric speech analysis.