Best feature selection for emotional speaker verification in i-vector representation

Abstract

This paper is dedicated to the gender-dependent text-independent speaker verification from Slovak emotional speech. To investigate the best speaker verification performance different features were extracted in front-end processing, namely MFCC (Mel-Frequency Cepstral Coefficients), LPC (Linear Prediction Coefficients) and LPCC (Linear Prediction Cepstral Coefficients), and their mapping into low-dimensional vector of fixed length was performed following the principles of i-vector method. In evaluation process of i-vectors scoring following Mahalanobis distance metric was employed.