Split learning (SL), a machine learning (ML) technique for collaborative training across devices and servers, partitions the model across entities to leverage computing power while preserving raw data privacy. However, device heterogeneity in terms of computation or communication capabilities causes the presence of stragglers, resulting in significant delays in the training process. This paper addresses the straggler problem in a wireless scenario with one edge server and multiple devices collaborating to train ML models using SL. We introduce a novel, low-complexity Coalition Formation Game (CFG) algorithm for SL in wireless networks. The CFG algorithm clusters work-ers based on their training times, effectively mitigating delays caused by stragglers. Specifically, this solution involves training devices negotiating to form or leave clusters to achieve better accuracy and/or shorter training delays, as well as selecting SL cut layers that best align with their communication and computing resources. Theoretical proof is provided, guaranteeing the termination of the CFG algorithm and the stability of the final clustering, where no devices have the incentive to leave the collaboration. We validate the proposed method on various datasets, and the results show that it strikes a good balance between lower training delay and high accuracy, consistently achieving top accuracy, and converging at the fastest speed.