Tree distributions approximation model for robust discrete speech recognition