Investigation of full-sequence training of deep belief networks for speech recognition