In recent years, Generative Adversarial Networks (GANs) have made significant progress in the field of speech enhancement. However, due to the high training difficulty of GAN, the adoption of complex network structures and a large number of parameters in mask estimation networks results in high computational costs and algorithmic delays. This paper proposes a lightweight model, MRGAN, based on Dense Temporal-Spectral Attention GAN networks for speech enhancement. MRGAN employs a dual-window, non-overlapping short-time Fourier transform to reduce algorithmic delays. The generator utilizes time-frequency conformer blocks for acoustic encoding, while the decoder predicts magnitude masks and phase information in a decoupled manner. The discriminator is metric-driven, mimicking PESQ scores as part of the loss function. Experimental results demonstrate that the proposed model achieves superior performance with fewer model parameters and lower algorithmic delays compared to existing models.