Privacy-preserving speaker recognition with cohort score normalisation