Abstract: The wavelet transform theory is introduced into the adaptive speech signal noise cancellation system.The principle of LMS adaptive speech signal noise cancellation algorithm with multiple-scale wavelet decomposition is analyzed.The MSWD-LMS algorithm converges faster than the classical LMS algorithm，and the spectrum dynamic range of auto-correlation matrix of the input vector which is decomposed to multiple-scale space is decreased.The variable step size LMS algorithm and the multiple-scale wavelet decomposition are merged into the adaptive speech signal noise cancellation system.A novel method is proposed，that is，a LMS adaptive algorithm with Variable Step Size based on Multiple-Scale Wavelet Decomposition（MSWD-VSS-LMS）.The novel algorithm can decrease the conditions of auto-correlation matrix of the input vector，as well as resolve the contradiction that fixed step size cannot result in fast convergence speed and low residual error simultaneously.Experimental results demonstrate that the new algorithm has not only better convergence property but also better stable mean square errors than the former algorithms.