Optimized Discriminative Kernel for SVM Scoring and Its Application to Speaker Verification