Variance compensation within the MLLR framework