Starting from the characteristic of voice signal itself, a new variable step-size LMS algorithm based on DCT is proposed in order to improve the comprehensive performance in adaptive noise cancellation, which combines the merits of normalized DCT-LMS algorithm and variable step-size LMS algorithm, and give full play to the decorrelation capability of DCT and rapid convergence effect of variable step-size algorithm. The c simulation results show that the new algorithm has faster convergence rate and smaller steady state error compared with traditional LMS and NLMS algorithm, however, the computational complexity is comparable to NLMS algorithm at the same time.