This paper deals with the problem of robust nonlinear feature transformation of speech signal. Starting with a conventional signal analysis a traina-ble nonlinear mapping based on statistical density modeling is proposed. The statistical model was represented by the parametric Gaussian distribution as well as by a nonparameteric histogram-based approach. The result of the transformation is a vector containing the a-posteriori-probabilities of a set of selected phonemic categories. Experiments were carried out using fluently spoken sentences from the German PHONDAT-I database.