Abstract: Based on the approximate sparsity of speech signal on the discrete cosine transform domain, we apply compressed sensing technology to speech enhancement area, and present a novel speech enhancement algorithm based on CS, which uses row echelon matrix to compress the noisy speech because of the robustness of row echelon matrix. Then, the original speech signal is reconstructed by using the modified orthogonal matching pursuit(OMP). Finally, We add a low-pass filter to smooth the reconstructed speech to achieve speech enhancement. The simulation results show that the proposed algorithm has a high speech enhancement capability, which not only improves output signal-to-noise ratio(SNR) and shortens reconstruction time, but also performs high robustness under various SNR.