Zero-shot learning (ZSL) aims to recognize objects of target classes by transferring knowledge from source classes through the semantic embeddings bridging. However, ZSL focuses the recognition only on unseen classes, which is unreasonable in realistic scenarios. A more reasonable way is to recognize new samples on combined domains, namely Generalized Zero Shot Learning (GZSL). Due to the fact that the source domain and target domain are disjoint and have unrelated classes potentially, ZSL and GZSL often suffer from the problem of projection domain shift. Besides, some semantic embeddings of prototypes are very similar, which makes the recognition less discriminative. To circumvent these issues, in this paper, we propose a novel method, called Learning Discriminative Domain-Invariant Prototypes (DDIP). In DDIP, both target and source domains are combined and projected into a hyper-spherical space, which is automatically learned by a regularized dictionary learning. In addition, an orthogonal constraint is employed to the latent hyper-spherical space to ensure all the class prototypes, including seen classes and unseen classes, to be orthogonal to each other to make them more discriminative. Extensive experiments on four popular benchmark and a large-scale datasets are conducted on both GZSL and standard ZSL settings, and the results show that our DDIP can outperform the state-of-the-art methods.