H- and C-level WFST-based large vocabulary continuous speech recognition on Graphics Processing Units