A signal detection algorithm adopting the hybrid belief propagation and expectation propagation (EP-BP) was proposed, by combining the orthogonal time frequency space (OTFS) technology and the generalized spatial modulation (GSM). The discrete probability distribution was projected into a multivariate complex Gaussian function, and the beliefs of the GSM transmission symbols were calculated via the iterative propagation of the mean vectors and the covariance matrices. A two-stage TS-EP-BP signal detection algorithm was designed to reduce the computational complexity. The activated antenna combination was determined via the EP-BP in the first stage, and in the second stage, the vector variable node (VN) in the factor graph was decomposed into multiple sub-VNs, and the invalid sub-VNs were cut out. Considering that the modulation symbols were independent to the activated antennas, the computation times of symbol belief were greatly reduced by the approximation of the univariate complex Gaussian distribution function. Bit error rates of different algorithms in different conditions were compared through simulation, and the results showed that, the proposed EP-BP and TS-EP-BP algorithms had better performance of bit error rate, especially, by the TS-EP-BP algorithm, a flexible performance-complexity tradeoff was striked by changing the maximum iteration times in the first stage.