This paper aims to provide a novel design of a multiscale framelets convolution for spectral graph neural networks. In the spectral paradigm, spectral GNNs improve graph learning task performance via proposing various spectral filters in spectral domain to capture both global and local graph structure information. Although the existing spectral approaches show superior performance in some graphs, they suffer from lack of flexibility and being fragile when graph information are incomplete or perturbated. Our new framelets convolution incorporates the filtering func-tions directly designed in the spectral domain to overcome these limitations. The proposed convolution shows a great flexibility in cutting-off spectral information and effectively mitigate the negative effect of noisy graph signals. Besides, to exploit the heterogeneity in real-world graph data, the heterogeneous graph neural network with our new framelet convolution provides a solution for embedding the intrinsic topological information of meta-path with a multi-level graph analysis.Extensive experiments have been conducted on real-world heterogeneous graphs and homogeneous graphs under settings with noisy node features and superior performance results are achieved.