An enhanced Fishervoice subspace framework for text-independent speaker verification

In this paper, we propose an enhancement to Fishervoice approach for speaker verification. In this framework, we first represent each utterance with a fix-length super-vector using Joint Factor Analysis (JFA) Gaussian Mixture Models (GMM). Multiple discriminant projections are then used on partitioned vectors of the super-vector for fast and effective matching. Experiments are presented on the core test of NIST SRE 2008 male corpora. We achieve 10.3% relative improvement on EER compared to the state-of-art JFA. Finally, by integrating the enhanced Fishervoice model and the JFA, the matching performance can be further improved. This demonstrates the effectiveness of the proposed framework.