Thanks to the development of deep learning, voice-visual cross-modal retrieval has made remarkable progress in recent years. However, there still exist some bottlenecks: How to establish effective correlation between voices and images to improve the retrieval precision and how to reduce data storage and speed up retrieval in large-scale crossmodal data. In this paper, we propose a novel Voice-Visual Cross-Modal Hashing (V2CMH) method, which can generate hash codes with low storage memory and fast retrieval properties. Specially, the proposed V2CMH method can leverage deep feature similarity to establish the semantic relationship between voices and images. In addition, for hash codes learning, our method attempts to preserve the semantic similarity of binary codes and reduce the information loss of binary codes generation. Experiments illustrate that V2CMH algorithm can achieve better retrieval performance than other state-of-the-art cross-modal retrieval algorithms. © Springer Nature Switzerland AG 2019.