Minimum classification error /eigenvoices training for speaker identification