Increasing robustness to training-test mismatch in speaker verification through shallow source modelling | IEEE Conference Publication | IEEE Xplore