Generative calibration models have shown to be an effective alternative to traditional discriminative score calibration techniques, such as Logistic Regression (LogReg). Provided that the score distribution assumptions are sufficiently accurate, generative approaches not only have similar or better performance with respect to LogReg, but also allow for unsupervised or semi-supervised training.

Recently, we have proposed non-Gaussian linear calibration models able to overcome the limitations of Gaussian approaches. Although these models allow for better characterization of score distributions, they still require the target and non-target distributions to be reciprocally symmetric.

In this work we further extend these models to cover asymmetric score distributions, as to improve calibration for both supervised and unsupervised scenarios. The improvements have been assessed on NIST SRE 2010 telephone data.