End-to-End Losses Based on Speaker Basis Vectors and All-Speaker Hard Negative Mining for Speaker Verification.