generate G.fst for each language model (tglarge from text, french-small from LIUM, and LM mix of both)