Noise compensation in i-vector space using linear regression for robust speaker verification

In recent years, i-vector has become the state-of-the-art input feature for speaker verification (SV) systems. Improving the noise robustness of i-vector based SV systems is of great research interest. Recently, maximum a posteriori (MAP) estimation of additive noise in i-vector space (i-MAP) has been reported to be effective for improving the noise robustness of SV systems. For i-MAP to be effective, the distribution of the feature vectors needs to be Gaussian. After applying the i-MAP algorithm, the resultant i-vectors are no more Gaussian distributed. This limits the iterative use of i-MAP for improved performance. In this paper, we explore how the performance of the i-MAP based noise compensation technique can be further improved. We use linear regression, for mapping the input noisy i-vector to a clean i-vector space. This speaker specific, regression based “cleaning” method, helped to further reduce the effect of additive noise on i-MAP cleaned i-vectors. In this approach, we used five consecutive elements of noisy i-vector to predict each element of the cleaned i-vector. Baseline system for all our experiments is i-vectors with support vector machine backend classifier (i-SVM). Experiments were performed using the female part of core short2-short3 trials of NIST 2008 speaker recognition evaluation (SRE). We obtained a relative performance improvement of 30% in Equal Error Rate (EER) over the non-cleaned i-SVM system using the i-MAP algorithm. Regression cleaning helped to achieve 12% relative improvement in EER for 0 dB noise level over the i-MAP i-SVM system.