The main purpose of our research is to build an automatic audio-visual speech recognizer. It is important to note that the corpus used is limited compared to other databases usually used to test speech recognizers. As we will explain further in this report, the system has to identify an item out of only 18 candidates, so that the test conditions are rather simple when compared to a recognition problem. Nevertheless, our goal is to study the benefit of an automatic lipreading on automatic speech recognition. In that sense, any training/test paradigm will be as informative as another. This is why we have first chosen a simple procedure.