Point cloud processing is a challenging task due to its sparsity andirregularity. Prior works introduce delicate designs on either local featureaggregator or global geometric architecture, but few combine both advantages.We propose Dual-Scale Point Cloud Recognition with High-frequency Fusion(DSPoint) to extract local-global features by concurrently operating on voxelsand points. We reverse the conventional design of applying convolution onvoxels and attention to points. Specifically, we disentangle point featuresthrough channel dimension for dual-scale processing: one by point-wiseconvolution for fine-grained geometry parsing, the other by voxel-wise globalattention for long-range structural exploration. We design a co-attentionfusion module for feature alignment to blend local-global modalities, whichconducts inter-scale cross-modality interaction by communicating high-frequencycoordinates information. Experiments and ablations on widely-adoptedModelNet40, ShapeNet, and S3DIS demonstrate the state-of-the-art performance ofour DSPoint.