Most of the applications can be organised as a sequence of dialogue phases where each phase needs a specific lexicon. For example in a menu-driven dialogue system each menu offers a list of choices with the corresponding commands: a subset of recognition vocabulary that is active at one time. It is clear that if the active vocabulary is reduced to these commands the confusion between words will be reduced as well as the time devoted to the speech comparison, while increasing the recognition rate. For example if the system requests an input that should be YES or NO within a complete lexicon of 100 words then there is no need to compare the input to all the 100 words. Nevertheless one has to consider the alternatives, synonyms, and the possible extra-commands that may be pronounced by the user (such as No thanks, Yes please, Operator, Stop, that's all, ...). The branching factor is a ``divide to conquer'' approach [Kinsey (1994)] that indicates that the technology supports vocabulary subsets. This may be achieved automatically if the dialogue manager or another integrated tool allows it or it may have to be done at an early stage either by the technology provider (e.g. HMM modelling of classes of words) or given by the application developer as a list of sublexica to activate each time the speech recogniser is on.