Abstract: Audio-visual alignment using video data is a conventional approach for the self-supervision of multi-modal representation learning. Nevertheless, the presence of background music, external ...