A new discriminative learning algorithm for predictive speech recognition is proposed. Based on the hidden control neural network (HCNN) [2] we use the patterns of a certain word class to train both the correct word model in the classical positive way and competing word models in a negative direction. Hence the training is performed in a discriminative way. The influence of the proposed algorithms was tested in several experiments using isolated digits. When using the new untraining algorithm the recognition rates could be increased significantly (from 87.9% to 93.4%). We compared the results of the advanced HCNN algorithm to other connectionist recognizers. With the pure MLP [7] we reached 91.1%, with the TDNN [9] we achieved recognition rates up to 94.3%.