Reconstructing Speech from Frame Differences