This paper focuses on modeling vocal command languages for a connected speech recognition system. The structure for representing the vocabulary and the syntax are very complex to be directly created by a user, so an easy, interactive way has been implemented for modeling command language of applications, and an automatic, transparent procedure gets the task of creating structures required by training and recognition algorithms. In the latter phase, the two major problems the system has to solve are the elimination of recursive structures (very frequent specially in office automation applications), and the need to break some sequences of words in order to cope with keyboard- speech mixed inputs in applications which cannot be completely interfaced by voice. For completeness, a description of how recognition uses the syntax has been provided.