Optimizing Multi-Taper Features for Deep Speaker Verification