The application of a speech dialog system is modeled as a hierarchy of sub-tasks in order to - have the possibility of structuring complex dialogs into sub-dialogs and thus restricting the lexicon to be recognized at each dialog state, - to model possible sequences of sub-dialogs, - to enable several instances of the same sub-tasks, - to make references and/or switches to prior discussed sub-tasks possible. Thus, also more complex applications can be modeled and handled without losing speech recognition accuracy or getting lost in different dialog themes. As these restrictions are quite natural users will not perceive any limitation.