Towards robustness in speaker verification: enhancement and adaptation

In this paper, we study the performance limits of a standard HMM speaker verification (SV) system in "adverse conditions" context using two real noises with different SNR level. Two different techniques are investigated : a) Adaptive noise cancellation represents one such potentially effective technique and refers to a class of adaptive enhancement algorithms based on the availability of a primary input source and a secondary reference source. b) Maximum Likelihood Linear Regression (MLLR) transformation which adapts the system to the new environment based on modifying the continuous densities of the HMM mixtures. We apply the MLLR to perform environmental compensation by reducing a mismatch due to additive noise effects. Experiments on 60 speakers Yoho corpus corrupted by different type of noises showed that the performance of the SV can approach the optimal performance with the noise cancelling technique and improves it by more than 6% with the MLLR method.