Partial AUC optimization based deep speaker embeddings with class-center learning for text-independent speaker verification.