OK, a couple suggestions. The first is to make sure that you are entering those items into LanguageModelGenerator as phrases rather than discrete words, which should raise the probability of them being detected as phrases. But in order to absolutely require that they are detected as phrases, you can try one of the following two approaches:
1. The not-so-easy but correct way: use a JSGF grammar instead of an ARPA model. This allows you to set rules for what can be recognized. It’s not-so-easy because you have to learn about how JSGF works and define the rule structure yourself.
2. The easy and cheap way: turn them into a single word (“MillerStreet”) and then re-insert the space for display to the user.