Maximum entropy discrimination (MED) feature subset selection for speech recognition