Sixth International Workshop on Models and Analysis of Vocal Emissions for Biomedical Applications (MAVEBA 2009)

Florence, Italy
December 14-16, 2009

Principal Component Analysis for HMM-Based Pathological Voice Detection

M. Sarria-Paja (1), G. Castellanos-Domínguez (2), N. Gaviria-Gómez (3)

(1) Intelligent Machines and Pattern Recognition Group, Instituto Tecnológico Metropolitano, Medellín, Colombia
(2) Control and Digital Signal Processing Group, Universidad Nacional de Colombia, Manizales, Colombia
(3) Facultad de Ingeniería, Universidad de Antioquia, Medellín, Colombia

This paper presents a methodology for feature selection in dynamic problems based on the analysis of the variation of linear components in acoustic features combined with an estimation of the ratio between a compactness measure to the separation measure. The methodology is applied to the automatic detection of voice disorders by means of stochastic dynamic models: results showed a significant reduction in the number of features, 96.6% of accuracy, and a 62.2% of computational cost reduction.

Index Terms. Dynamic features. HMM, PCA, feature selection, pathological voice, clustering

Full Paper (reprinted with permission from Firenze University Press)

Bibliographic reference.  Sarria-Paja, M. / Castellanos-Domínguez, G. / Gaviria-Gómez, N. (2009): "Principal component analysis for HMM-based pathological voice detection", In MAVEBA-2009, 69-72.