For many applications in cursive script recognition the vocabulary is restricted to a small and fixed set of words. In a recognition approach such as Hidden Markov Models, for each of these words a model is constructed and trained. In the recognition task an unknown pattern is matched with all of these models to find the most likely class. In our paper, we describe a method for reducing the size of vocabulary depending on the actual input. In contrast to many other techniques, our approach is not using any topological features. The reduction system is directly based on the quantized feature vectors which are used as input for the HMMs. Thus very little additional work is required for lexicon reduction. The proposed approach was successfully tested on two different systems with small lexicons. In both cases, the lexicon could be reduced to 25% of its original size without increasing in the error rate.