Speech denoising focuses on removing or reducing noise from an audio signal to improve the quality of the resulting speech in real-world applications such as automatic speech recognition (ASR), virtual assistants, or telephone communication systems. In this paper, we propose the Attention Gate-based U-Net architecture for the speech enhancement task, named AGSEU-Net. The attention gate serves a pivotal function by prioritizing informative regions within the spectrogram, thereby facilitating more efficient training and enhancing the model’s capacity to generalize across varying acoustic features. AGSEU-Net introduces a novel spatial attention strategy specifically tailored to improve spectrogram representation. Through the dynamic selection of salient areas associated with target structures of varying geometries, the attention mechanism directs the network’s focus toward the most informative content. When embedded within the U-Net framework, this attention module effectively suppresses non-essential regions and accentuates key spectral components essential for robust speech enhancement. Empirical evaluations demonstrate that the proposed model outperforms both the baseline and prior approaches in terms of noise suppression and speech quality improvement.