In this paper, we introduce a novel temporal convolutional shrinkage network to enhance feature learning from noisy speech signals. Taking into account the non-stationary nature of speech signals, we introduce an approach that integrates time-varying soft thresholding with a temporal convolutional network. This enhancement aims to improve the robustness of the KWS model against noise. Our experiments demonstrate the effectiveness of the proposed model in noise suppression, resulting in an improved performance of the KWS system in noisy environments. Furthermore, an ablation study provides verification of the efficacy of the proposed shrinkage layer and the soft thresholding processing.