This paper presents a method for extracting discriminative key poses for skeleton-based action recognition. Poses are represented by normalized joint locations, velocities and accelerations of skeleton joints. An extended label consistent K-SVD (ELC-KSVD) algorithm is proposed for learning the common and action-specific dictionaries. Discriminative key poses are represented by the atoms of the action-specific dictionaries. With the specific dictionaries, sparse codes are obtained for representing action instances through max pooling and temporal pyramid. A SVM classifier is trained for action recognition. The proposed method was evaluated on the MSRC-12 gesture and MSR-Action 3D datasets. Experimental results have shown that the proposed method is effective in extracting discriminative key poses.