Zero-shot learning (ZSL) is a transfer learning paradigm that aims to recognize unseen categories just by having a high-level description of them. While deep learning has greatly pushed the limits of ZSL for object classification, ZSL for gesture recognition (ZSGL) remains largely unexplored. Previous attempts to address ZSGL were focused on the creation of gesture attributes and algorithmic improvements, and there is little or no research concerned with feature selection for ZSGL. It is indisputable that deep learning has obviated the need for feature engineering for problems with large datasets. However, when the data are scarce, it is critical to leverage the domain information to create discriminative input features. The main goal of this work is to study the effect of three different feature extraction techniques (, and features) on the performance of ZSGL. In addition, we propose a bilinear auto-encoder approach, referred to as Joint Semantic Encoder (JSE), for ZSGL that jointly minimizes the reconstruction, semantic and classification losses. We conducted extensive experiments to compare and contrast the feature extraction techniques and to evaluate the performance of JSE with respect to existing ZSL methods. For classification scenario, irrespective of the feature type, results showed that JSE outperforms other approaches by 5% (<0.01). When JSE is trained with features in condition, we showed that JSE significantly outperforms other methods by 5% (<0.01)).