The deep network model lack the ability to express the underlying feat ures when used to denoise speech end-to-end and only use convolution causing p ar ameter redundancy.Inspired by these,a generative adversarial network with the residual structure and attention mechanism is proposed (Attention Res-UNetGAN). The speech is denoised in the raw waveform domain.The generator network is U-N e t including downsampling layer,middle layer and upsampling layer.The middle la yer is an improved Residual-Residual Block (RRB),which is formed by the residu a l splicing of two basic residual blocks.Attention skip and direct skip are used between symmetrical downsampling layer and upsampling layer.Denoising on the V CTK (Voice Bank corpus) and evaluated by six objective speech quality evaluation indicators such as PESQ,Attention Res-UNetGAN has increased by an average of 9.13% compared with the generative adversarial network.