Fast scoring for mixture of PLDA in i-vector/PLDA speaker verification

With the ubiquitous of mobile phones, users of speaker verification systems will perform authentication anywhere at anytime. As a result, practical speaker verification systems need to deal with utterances of different noise levels. Recently, an SNR-dependent mixture of PLDA model was proposed to deal with such practical situation. However, the scoring function of this model is significantly more complex than the conventional one. This paper proposes a method to reduce the computation burden of this mixture PLDA model. The idea is based on the observation that for most utterances, the posterior probabilities of SNR are very sparse so that it is possible to consider the top Gaussian only during scoring. The method effectively reduces the computational complexity from O(K2D3) to O(D3), where K and D are the number of mixtures and i-vector dimension, respectively. Experimental results based on NIST 2012 SRE suggest that the proposed method can reduce computation time by 60% with very minor degradation in performance.