A new proactive defense against partial speech manipulation via self-embedding steganography is proposed, which can detect and restore partially deepfaked audio without any training. This method embeds a clean speech signal with a compressed representation of itself, enabling post-hoc extraction of reference content.