Abstract: There is a certain redundancy in the primary function of wavelet for dyadic wavelet transform. So, the de-noising result by dyadic wavelet transform is better than that by discrete wavelet transform. How to estimate the noise threshold and select the threshold function exactly will affect obviously the de-nosing accuracy. Analyzing the characteristics of the dyadic wavelet transform of Gaussian noise, a new approach to signal de-noising is proposed to improve the dyadic wavelet transform by introducing a new threshold to get rid of the constant error when using Donoho's soft threshold to estimated and decompose the wavelet coefficients. Simulation results indicate that the new approach will not only suppress effectively the pseudo-Gibbs phenomena at the singular point of signal waveform, but provide a higher de-noising accuracy than Donoho's hard and soft-threshold methods.