In recent years, speech recognition researchers have proposed the use of Gaussian warping as a step in the computation of input speech feature parameters. This warping is intended to reduce the mismatch between the actual statistical distribution of parameters and the distribution hypothesized in the acoustic models used, i.e., the Gaussian distribution. In this paper, we compare variants of Gaussianization, including off-line and windowed (short-term) versions, which we evaluate on a corpus of Canadian Parliamentary Debates.