In this paper we propose a way to include a phonetic representation using binary distinctive phonetic features into the probabilistic framework of a hybrid system for word recognition. This could be done under two assumptions: 1.) The features are changing synchronously at the borders of the phonemes and 2.) the features are conditionally independent with one another. We report experiments we have done with a hybrid system working with a feature representation with which we get a recognition rate of 96.9 % word recognition on a spelled letter task. Then we describe some ways to weaken the assumption of independence between the features. We demonstrate that these ways do not improve significantly the recognition rate but lead to a simpler system with the same performance.