Effective background data selection in SVM speaker recognition for unseen test environment: More is not always better