Abstract: Aimed at the problem that residual neural network has too many hyperparameters in speech deception detection, and the high-frequency features are not prominent enough, a ResNeXt-Attention network (RA-Net) fused with attention mechanism is proposed. RA-Net used residuals combined with grouped convolution, replaced large convolution kernels with a set of small convolution kernels, and used MFM (max feature map) as a new splicing method. The added attention mechanism reduced the attention to edge information by learning the original feature information. Experiments on the ASVspoof2019 data set show that compared with the baseline Gaussian mixture model (GMM), the equal error rate (EER) of RA-Net is reduced by 4.72 percentage points and 6.23 percentage points. And the EER is reduced by 0.69 percentage points and 0.89 percentage points compared with the residual network (ResNet). The validity of the model is confirmed.