Abstract: The source signal recovery model for underdetermined blind source separation based on compressed sensing(CS) is constructed, and the recovery effect of three algorithms separately based on the complementary matching pursuit(CMP), the L1 based complementary matching pursuit(L1CMP) and modified Newton radial basis function(NRASR) are compared by simulation. Results show that as to the completely sparse source signals in time domain, the recovery effect of the three algorithms are similar, while the calculation complexity of L1CMP is the lowest. As to the completely sparse source signals in transformation domain, the recovery effects of CMP and L1CMP are similar, but that of NRASR is worse. When the source signals are incompletely sparse in time domain, the recovery effect of CMP is worse, and those of L1CMP and NRASR are similar. So based on comprehensive consideration, the L1CMP algorithm is the best in the three algorithms. As to the case of the source signal number and observation signal number are small, the recovery effect would decline in time domain. The sparse representation method combined with the CS reconstruction algorithms can get good source signal recovery effect.