Abstract: Neurocognitive science research shows that human brain often combines face information on cross-modal interaction analysis during the speech perception process. Nevertheless, existing cross-modal face-voice association methods still face the various challenges such as sensitivity to complex samples, lack of supervised information and insufficient semantic correlation, which mainly due to the lack of mining common semantic embeddings. To tackle these problems, we present an efficient cross-modal face-voice matching method from bi-pseudo label based self-supervised learning. First of all, we introduce a cross-modal weighted residual network to learn face-voice common embeddings, and then propose a novel self-supervised learning method for bi-pseudo label association, which learns the latent semantic supervision of one modality to supervise the feature learning of another modality. Accordingly, based on this interactive cross-modal self-supervised learning, the highly correlated face-voice associations can be well learned. Besides, in order to increase the discrimination of mining supervised information, we further construct two auxiliary losses to make the face-voice features of the same samples closer, while pushing the features of different samples to be far away. After a large number of experiments, the innovative method proposed in this paper has achieved a comprehensive improvement in the cross-modal face-voice matching task compared with the existing work.