In this letter, two sphere-decoding (SD) algorithms for spatial modulation (SM) are proposed to reduce the computational complexity of maximum-likelihood (ML) optimum detector. Through theoretical analysis and simulation results, it is shown that the proposed algorithms offer a near-optimum performance. Under the same spectral efficiency, both the SDs offer a significant reduction in computational complexity with respect to not only the ML detector but also the conventional SDs for SM. Furthermore, it is shown that one of the proposed SDs is always superior to the others in computational complexity, and the other one always offers the same bit error ratio (BER) as ML-optimum detection.