In Hidden Markov models, speech features  are modeled by Gaussian
distributions. In this paper, we propose to gaussianize the features to
better fit to this modeling. A distribution of the data is estimated and
a transform function is derived. We have tested  two  methods of the transform
estimation (global and  speaker based). The results are reported on
recognition of isolated Czech words (SpeechDat-E) with
CI and CD models and on medium vocabulary continuous speech
recognition task (SPINE). Gaussianized data provided in all three
cases results superior to standard MFC coefficients proving, that the
gaussianization is a  cheap way to increase the recognition accuracy