Using Fishervoice to enhance the performance of I-vector based Speaker Verification System

I-vector is a popular feature representation technique in speaker verification systems. In this paper, we use Fishervoice algorithm in combination with i-vector feature representation to improve speaker verification performance. By applying the Fishervoice model to map the i-vector into a low-dimensional discriminant subspace, the intra-speaker variability can be reduced and the discriminative class boundary information can be emphasized for enhanced recognition performance. Experiments on NIST SRE 2008 core test task show that the proposed framework achieves 19.9% and 8.5% dramatic relative decrease in EER and minDCF metrics respectively compared to the state-of-the-art PLDA based method.