A little bit more info about the internals of the recognition engine would be welcome. It seems that one needs to manually supply sets of keywords and key phrases?