Hybrid threshold approach in text-independent speaker verification