We have known from the past research based on the LMS algorithm that normazied LMS (NLMS) algorithm enlarges the dynamic range of input, but it is highly sensitive to the noise disturbance. Variable step-size LMS algorithm (VSSLMS) which uses auto-correlation of errors can provide fast convergence, eliminate the influence of independent noise, thus give good performance in non-stationary environment. But it can not perform well with input which has large dynamic range. This paper combinates their advantages, and the new algorithm can perform well both in the environment of low SNR and the environment with large dynamic range of input.