Speaker weight estimation from speech signals using a fusion of the i-vector and NFA frameworks

In this paper, a novel approach for automatic speaker weightestimation from spontaneous telephone speech signals is proposed. In thismethod, each utterance is modeled using the i-vector framework which isbased on the factor analysis on Gaussian Mixture Model (GMM) meansupervectors, and the Non-negative Factor Analysis (NFA) frameworkwhich is based on a constrained factor analysis on GMM weights. Then, theavailable information in both Gaussian means and Gaussian weights isexploited through a feature-level fusion of the i-vectors and the NFAvectors. Finally, a least-squares support vector regression (LS-SVR) isemployed to estimate the weight of speakers from given utterances.The proposed approach is evaluated on the telephone speech signals ofNational Institute of Standards and Technology (NIST) 2008 and 2010Speaker Recognition Evaluation (SRE) corpora. Experimental results over2339 utterances show that the correlation coefficients between actual andestimated weights of male and female speakers are 0.56 and 0.49,respectively, which indicate the effectiveness of the proposed method inspeaker weight estimation.

Jaar van publicatie:2015