A lot of research aims to improve accuracy in end-to-end speech recognition, and achieves higher accuracy on various famous corpora. However, there are many languages which do not have enough data to build their speech recognition system in the world. The system often cannot get a reliable result and be used in the real-world. Therefore, how to build a robust low-resource speech recognition system is an important issue in speech recognition. In this paper, we use ESPnet toolkit to implement an end-to-end speech recognition model based on sequence-to-sequence architecture, and use Fairseq toolkit to implement an unsupervised pre-training model for assisted speech recognition. In addition, we use unlabeled speech data to help extract speech features, and transfer a speech recognition model with sufficient corpus to Hakka speech recognition with less corpus through transfer learning. Experimental results show that we establish a more robust low-resource Hakka speech recognition system.