The number of words that an engine is able to recognise is a large factor in its recognition accuracy. The smaller the vocabulary the more probable the engine is to pick the correct word, for the simple reason that it has fewer choices.
As vocabulary increases the user is given greater freedom and an increasingly natural interface, but the complexity of computation and recognition inaccuracies both increase as a result.
As a vocabulary increases in size, the number of similar sounding words will also rise. This will in turn lead to more frequent inaccuracies in recognition.
Since all-encompassing vocabularies are not feasible, a general purpose speech recognition engine should therefore be capable of using smaller vocabularies on an application to application basis.