Abstract: Seismic phase picking is a critical task in earthquake monitoring, as its accuracy directly impacts the precision of hypocenter localization and magnitude estimation. However, traditional methods are often limited in their ability to capture the characteristics of complex seismic signals. This study proposes a dual-branch deep learning model that integrates a multi-scale attention mechanism and short-time Fourier transform (STFT). The model extracts temporal features through a time-domain branch and captures time-frequency representations via a frequency-domain branch, while leveraging the attention mechanism to enhance multi-scale features. Experimental results show that within a 100 ms error threshold, the proposed model achieves a P-wave picking precision and recall of 95.69% and 88.97%, and an S-wave precision and recall of 87.98% and 77.25%, respectively. The mean and standard deviation of arrival time error for the P-wave are 18.76 ms and 27.13 ms, while for the S-wave they are 25.97 ms and 36.14 ms. Moreover, the model contains only 0.35M parameters and incurs a computational cost of 71.38M FLOPs. Compared with existing models, the SEN model not only achieves competitive performance but also demonstrates advantages in model size and computational efficiency, offering strong potential for real-time seismic monitoring applications.