In this paper, we propose a novel high-level action representation using joint spatial-temporal attention model, with application to video-based human action recognition. Specifically, to extract robust motion representations of videos, a new spatial attention module based on 3D convolution is proposed, which can pay attention to the salient parts of the spatial areas. For better dealing with long-duration videos, a new bidirectional LSTM based temporal attention module is introduced, which aims to focus on the key video cubes instead of the key video frames of a given video. The spatial-temporal attention network can be jointly trained via a two-stage strategy, which enables us to simultaneously explore the correla- tion both in spatial and temporal domain. Experimental results on benchmark action recognition datasets demonstrate the effectiveness of our network.