This study proposes a cross-modal learning framework that learns modality-agnostic embeddings for six modalities: language, temporal 2D/3D skeletons, IMU, GPS, and Wi-Fi. Existing approaches such as ImageBind align each target modality with images. Consequently, they are restricted to modalities for which paired image data can be prepared a priori, which makes them unsuitable for time-series tasks, such as action recognition. To overcome this limitation, we propose "chained modal alignment," which aligns the six modalities sequentially via temporal 3D skeletons, thereby enabling emergent zero-shot action recognition based on a shared embedding space. Because the proposed method relaxes the requirement of direct pairing with images, it supports a wider range of modality combinations and thus expands practical applications. In addition, we introduce "chained modal transfer learning," which transfers the knowledge learned between language and 3D skeletons to other modalities with data that cannot be directly paired, thereby improving generalization. Experiments on three public datasets spanning all six modalities quantitatively validate the effectiveness of the proposed method for emergent zero-shot action recognition.