Modified Segmental Histogram Equalization for robust speaker verification

It is well known that when there is an acoustic mismatch between the speech obtained during training and testing the accuracy of speaker recognition systems drastically deteriorates. In this paper we propose Modified Segmental Histogram Equalization to improve the robustness of a speaker verification system operating in telephone environments. The technique transforms the features extracted from short adjacent segments of speech within an utterance such that their statistics conform to that of a Gaussian distribution with zero mean and unity variance across all recording conditions. In doing so, the feature statistics become less environment-dependent. Experiments performed on the NIST 2000 database show significant improvements in performance.