Applying feature reduction analysis to a PPRLM-multiple Gaussian language identification system

Abstract

This paper presents the application of a feature selection technique such as LDA to a language identification (LID) system. The baseline system consists of a PPRLM module followed by a multiple-Gaussian classifier. This classifier makes use of acoustic scores and duration features of each input utterance. We applied a dimension reduction of the feature space in order to achieve a faster and easier-trainable system. We imputed missing values of our vectors before projecting them on the new space. Our experiments show a very low performance reduction due to the dimension reduction approach. Using a single dimension projection the error rates we have obtained are about 8.73% taking into account the 22 most significant features.