通过一些额外的语音对训练好的说话人模型进行测试统计出其输出评分的均值和方差（u,δ）当真正进行测试时，对输出模型的分数score，用规整之后的score =（score-u）/δ代替。使得结果更加合理。-Speaker model trained by some additional voice test statistics of the mean and variance of its output rated (u, δ) when the real test, the fraction

Tnorm利用大量冒充者模型来得到均值u和方差δ，再用这两个数据带入到score =(score-u)/δ得到规整化之后的分数。这样使得结果更加合理。-Tnorm impostor model to get mean u and variance δ, then these two data into the score ' = (score-u)/δ score after regularization. This makes the results more reasonable.