Maximizing Distance between GMMs for Speaker Verification Using Particle Swarm Optimization
Minseok Kim, IL-Ho Yang, Ha-Jin Yu · 2008
In this paper, we propose a feature transformation method to maximize the distances between the Gaussian mixture models for speaker verification. The feature transformation matrix is optimized by using particle swarm optimization. We evaluate the transformation using YOHO speech data, and the transformation is applied to some speakers who give poor performance. As the result, the overall equal error rate is reduced to 1.71% from 1.97% of the baseline.