paper：Bayesian Speaker Verification with Heavy-Tailed Priors：
Note that these normalizations break the symmetry between D 1 and D 2 , so in experimenting with this question it is natural to seek a score normalization procedure which preserves this symmetry. The simplest choice is to normalize the score s of a the pair (D 1 ,D 2 ) using the formula
where the mean µ 1 and standard deviation σ 1 are calculated by matching D 1 against a given imposter cohort and similarly for µ 2 and σ 2 . We refer to this type of score normalization as s-norm. It has been our experience that, in situations where score normalization is helpful, s-norm is more effective than z-norm or zt-norm.
D1，D2分别是注册和测试中的某一个句子
paper：Comparison of speaker recognition approaches for real applications
The AS-Norm（adaptive s-norm） is derived from the AT-Norm [13], but preserves the symmetrical property of the S-Norm [6]. The matching score s of two i-vectors i 1 and i 2 is normalized according to
where µ 1 and σ 1 are the mean and standard deviation of the scores obtained by matching i 1 against a normalization subset N 2 depending of i 2 , and the same notation dually applies to the second term in parenthesis. The selection of the normalization subset follows the procedure in [13]. Our normalization set includes 273 male and 348 female segments, selected from different languages conversations of the NIST SRE 04/05/06. It is worth noting that we always performed gender dependent normalization.
paper：Unsupervised Speaker Adaptation based on the Cosine Similarity for Text-Independent Speaker Verification
讲的是Znorm Tnorm ZTnorm 以及Snorm，不过这个Snorm和上面的一样的，所以不再赘述，后期需要查阅论文