Rebuilding the Statistical Model with Cleaner Voice