We describe a novel feature compensation algorithm based on the minimum mean square error (MMSE) estimation and stereo training data for robust speech recognition. The proposed algorithm can be viewed as a piece-wise linear transformation between the noisy and clean feature spaces, where both spaces are modeled by means of vector quantization (VQ) codebooks. By means of this VQ modeling, we show that a very efficient estimator can be obtained in terms of computational cost and recognition accuracy. Also, two approaches are proposed in order to compensate the acoustic noise distortion. First, we propose a novel formulation for the normalization of noisy feature vec- tors. Second, a novel subregion-based modeling is applied to obtain a better representation of the differences between noisy and clean domains. The experimental results on noisy digit recognition show a relative improvement of 61.49% over the baseline when clean acoustic models are used. Furthermore, important improvements are achieved in comparison with other similar approaches.