Speech recognition systems with small and medium vocabulary are used as natural human interface in a variety of real world applications. Though they work well in a quiet environment, a significant loss in recognition performance can be observed in noise-contaminated real world applications. In order to make such a system more robust, the development of a neural network based noise reduction module is described in this paper. Using standard feedforward networks, several topologies have been tested to learn about the properties of neural noise reduction. For the develop ment of a sufficiently robust nonadaptive system, information about the characteristics of the noise and speech components of the input signal including context information was taken into account. This paper is focused on the stepwise experiment-oriented improvement of a basic linear neural noise reduction network. Evaluation of each step is done by a real world speech recognition task and includes experiments with changing input signal characteristics to test the robustness of this system. 0-7803-0559-0 192 $3.00 0 1992 IEEE