<p>In many speech-related applications, speech enhancement models are required to be lightweight. However, these lightweight models often suffer from serious performance degradation under adverse noise conditions. To address this issue, this paper proposes a lightweight Attention-Guided Network (ADNet) based on a U-Net architecture, aiming to enhance speech quality and robustness in complex acoustic environments. In this work, the term ADNet specifically refers to a lightweight attention-guided architecture designed for single-channel speech enhancement. Specifically, the noisy speech is first converted into time-frequency features and sent to an encoder, which consists of a set of lightweight MET Transformer blocks embedded within the U-Net structure. Then, the encoded features are further processed at the bottleneck, where Squeeze-and-Excitation Block (SEBlock) and Frequency-Dilated Attention Block (FDAB) are integrated to recalibrate channel responses and broaden the receptive field for more comprehensive contextual modeling. Afterward, Attention-Guided Skip Connection (A-Gate) modules are introduced in the decoder to adaptively weight and fuse encoder and decoder features, helping suppress residual noise propagation and contributing to improved perceptual quality. Finally, the enhanced magnitude and phase spectra are simultaneously estimated to reconstruct the time-domain waveform. Compared with lightweight baselines and representative models, ADNet achieves notable improvements in speech enhancement, particularly under low-SNR conditions, demonstrating the effectiveness and robustness of the proposed model.</p>