Reducción de espacios de entrenamiento dinámico en la identificación de disfonías

Abstract

A method for training spaces reduction based on the features dynamic extraction is developed. The proposed orthogonal transformation turns out to be effective in the principal components determination with greater influence in the time process variability information that is used for both the features reduction and features selection with greater discriminative capacity. In particular, the contours effectiveness of acoustic features varied is analyzed in the hidden Markov model training for the hipernasal voices identification. An effectiveness of until 95% with a reduced set of not more than 2-3 features is obtained. © Springer-Verlag Berlin Heidelberg 2007.