Is your question about how to dynamically create a JSGF/dictionary pair using LanguageModelGenerator? You can’t create a JSGF using LanguageModelGenerator, but you can put the words from an existing JSGF into LanguageModelGenerator and use the resulting .dic file along with your pre-existing JSGF. Then you just pass the two files into startListening: (remembering to set isJSGF to TRUE) instead of an ARPA file and a .dic file.
It isn’t heavily documented because I can’t provide much support for the fairly complex subject of the JSGF standard and the Sphinx implementation thereof, so it is more something I make available as an option than actively promote, if that makes sense.