By analysis of variable step size NLMS ( VSS-NLMS ) algorithms and compare their performances in far-talk and double-talk mode, an improvement is made on the NEW-NPVSS algorithm. The improved algorithm has better convergence in double-talk mode. Then a new variable step-size NLMS algorithm is put forward based on the filter coefficient gradient. When the filter coefficient of gradient is less than threshold value, fixed step-size is used in update the filter coefficients. Conversely, it stops updating the filter coefficients, and uses the coefficient in far-talk mode to replace the current coefficient. The simulation results show that the proposed algorithm has better convergence than the other variable step-size NLMS algorithms in far-talk mode, and has better convergence compared with the fixed step NLMS, SVSS in double-talk mode.