Feature selection in pathological voice classification using dinamyc of component analysis

Abstract

This paper presents a methodology for the reduction of the training space based on the analysis of the variation of the linear components of the acoustic features. The methodology is applied to the automatic detection of voice disorders by means of stochastic dynamic models. The acoustic features used to model the speech are: MFCC, HNR, GNE, NNE and the energy envelopes. The feature extraction is carried out by means of PCA, and classification is done using discrete and continuous HMMs. The results showed a direct relationship between the principal directions (feature weights) and the classification performance. The dynamic feature analysis by means of PCA reduces the dimension of the original feature space while the topological complexity of the dynamic classifier remains unchanged. The experiments were tested with Kay Elemetrics (DB1) and UPM (DB2) databases. Results showed 91% of accuracy with 30% of computational cost reduction for DB1.