Task-oriented dialogue systems depend on dialogue state tracking to keep track of the intentions of users in the course of conversations. Recent studies in dialogue state tracking have achieved good performance, although the great majority of them do not consider slot correlation and just predict the value of every slot separately. In this work, we propose an efficient slot correlation learning network that can capture the correlations among slots as precisely as possible. Specifically, a BERT-base-uncased encoder is first applied to encode the dialogue context, slot names and their corresponding values. Second, we design a cross multi-head attention module to calculate and fuse attention among dialogue context embedding, slot name embedding and corresponding value embedding, which extracts relevant features and provides them to other components to fully catch the slot-specific information of every slot. Finally, a transformer encoder module is used to catch the correlations among slots. Experimental results on MultiWOZ 2.0, MultiWOZ 2.1, and MultiWOZ 2.4 datasets demonstrate the effectiveness of our approach with 55.14%, 57.22% and 76.93% joint goal accuracy, respectively, which achieves new state-of-the-art performance.