Open-vocabulary semantic segmentation aims to assign each pixel a semantic label from an open set of categories. Recent advances in cost aggregation-based methods have demonstrated promising improvements for this task, and their core idea is to perform both spatial aggregation and class aggregation on the cost map to achieve precise segmentation results. In spatial aggregation, the Transformer has limitations in local pattern modeling, and the convolution operation has weaknesses in long-range dependency. In this work, we proposed a Dynamic Multi-scale Aggregation Network (DMSA-Net). The DMSA-Net incorporates two critical modules: the Multi-scale Feature Perception (MSFP) module that projects an initial cost map into high-dimensional embedding spaces to capture discriminative cross-modal correlations between textual and visual features, and the Dynamic Feature Aggregation (DFA) module that employs a Dynamic Multi-scale Convolution (DMSC) for localized feature refinement as well as a Global Efficient Attention (GEA) mechanism for long-range dependency modeling. The DMSC and GEA mechanisms are adaptively fused, enabling dynamic adjustment of local-global feature contributions during the aggregation process. Comprehensive evaluations on multiple datasets demonstrate the superiority of our DMSA-Net in open-vocabulary semantic segmentation. The code will be announced at https://github.com/SWU-CS-MediaLab/DMM-DMSA .