A unique "topic" should exist for each classification of document that an application can have. Example topics might be "E-mail", "Formal Writing", or "C-Code", three text-entry mediums where the style of language used is vastly different. Voice Dictation will store information on topics away so that any information the recognizer learns will be stored, along with any glossary entries that the user generates specifically for the topic.
Adding a new topic might take a lot of hard disk or memory, so unless an application has a very good reason for creating a new topic, it should use an existing topic. Topics cannot be merged. Much of the information stored within a topic data structures is engine dependent so switching dictation engine vendors may cause a loss of some data.