I-VECTOR KULLBACK-LEIBLER DIVISIVE NORMALIZATION FOR PLDA SPEAKER VERIFICATION

I-vector and Probabilistic Linear Discriminant Analysis (PLDA) represents the state-of-the-art in the speaker verification system. In PLDA, the i-vectors are assumed to follow Gaussian distribution. However, this assumption results in poor modeling without Gaussianization. Different from previous Gaussianization methods, in our proposed method, we make no restriction towards the original distribution of i-vectors for flexibility and universality. To optimize the Gaussian transformation function, Kullback-Leibler divergence (KLD) is introduced to measure the distance between the two distributions. By minimizing the KLD value under the development data, we can search out the optimal parameters in transformation function. The proposed method shows significant improvement on NIST SRE 2008 core set; together with length normalization (LN), a famous Gaussianization method, can further improve the verification accuracy.