Artificial Neural Networks have generated widespread research activity also within the speech recognition community. Although ANN has a great potential, traditional methods have so far outperformed ANN-based solutions for speech recognition in most cases. In this study, we propose a method for non-linearly compressing the data presented to the classifier. The non-linear compression reduces the amount of input data, thereby simplifying the task of the network which in turn leads to improved performance. In a isolated word recognition task, using an alpha-digit vocabulary, the proposed system obtained a recognition score of 99. 1%, better than any of the "traditional" speech recognizers tested on the same task.