<p>Speaker verification is a crucial technology for confirming speaker identity by exploiting unique characteristics in speech. With the advancement of deep learning, convolutional neural networks (CNNs), such as ResNet, have been widely adopted for this task. However, existing models often rely on simple concatenation or weighting for feature fusion, lacking effective multi-scale modeling capability. This limits their ability to simultaneously capture local time–frequency details and global statistical information. As a result, the discriminative power of the learned embeddings is reduced. To address this issue, we propose a lightweight and pluggable multi-scale residual module, termed Multi-Scale Gated Fusion (MSGF). The module employs depthwise separable convolutions (DSConv) to enlarge the receptive field and model long-term dependencies. During the fusion stage, a channel attention mechanism is incorporated. It explicitly emphasizes discriminative features while suppressing redundant information. Meanwhile, learnable residual gating parameters adaptively regulate the contribution of each branch. This enhances training stability. Experiments were conducted on the VoxCeleb1-O/E/H datasets. They demonstrate that the proposed model consistently outperforms the Enhanced Res2Net (ERes2Net) baseline. It achieves relative reductions of 6.4%, 5.8%, and 8.0% in Equal Error Rate (EER) and 20.6%, 6.0%, and 4.7% in Minimum Detection Cost Function (MinDCF), respectively. Meanwhile, the model maintains a lightweight design with a 2.1% reduction in parameters. These results demonstrate the effectiveness and practicality of MSGF, providing an efficient solution for speaker verification tasks.</p>