<p>Insulated gate bipolar transistor (IGBT) modules are core components of power electronic systems, and void defect in their welding layers severely compromise system reliability and service life. Accurate void segmentation is critical for quality control yet remains challenging due to complex morphological variations and low contrast. Moreover, highly accurate, real-time, and robust methods are crucial for industrial applications. To address these issues, we propose the global-local contextual attention network (GLCA-Net), a lightweight encoder-decoder architecture for real-time void segmentation in IGBT welding layers. In the encoder, to capture both global contextual information and fine-grained local details across multiple scales, we propose a global-local multi-scale attention (GLMSA) module. In the decoder, inspired by the skip connection of DeepLabV3+, we devise a scale-aware channel spatial attention (SCSA) module to take full advantage of the multi-scale features from the encoder. Validated on two in-house IGBT welding void defect datasets (IGBT-WVD-Chip and IGBT-WVD-DBC), GLCA-Net outperforms UNet, DeepLabV3+, TransUNet, SwinUNet and BiSeNet. It achieves 71.49% Dice and 56.07% IoU on the Chip dataset, and 66.43% Dice and 54.29% IoU on the DBC dataset, with merely 0.244 M parameters and 182 FPS inference speed. This work provides a technical reference for the automatic void detection of IGBT welding layers, thus facilitating the improvement of their manufacturing quality.</p>