Most of what I do with speech recognition centers around very limited grammars, designed to improve recognition accuracy in noisy environments. We also have a requirement to support XP, since most of our users are on Tablet PCs running XP SP2.
The latest issue of MSDN Magazine (not online yet) has an interesting article by Eric Brown on using the Text Services Framework to provide full-featured dictation in apps running on Vista. Among other features, TFS looks like an interesting approach to the problem of allowing your users to train the recognition engine as they dictate -- a feature which is of great interest to many developers, but which I have very little experience with at the point.