<p>Although offline handwritten character recognition has been explored for decades, accurate recognition of handwritten characters remains a challenging task. In this paper, a multiscale attention denoising diffusion probability model is proposed to solve the problem of low accuracy in multilingual offline handwriting character recognition caused by imbalanced samples. This model aims to expand the minority class samples to mitigate imbalance issues, thereby improving recognition accuracy. First, a multilingual offline handwritten character dataset consisting of 11 scripts and 4287 classes is constructed and published on the GitHub platform. Then, a traditional convolution extracts relatively shallow features from the original image, which cannot fully utilize complex features. ConvNeXt is introduced in the encoder section, and shallow and deep features are combined to enhance the feature extraction capability of the model. A dilated convolution module is added to the middle layers of the model to address the reduced image resolution issue caused by sampling, while maintaining a high image resolution and expanding the receptive field. A multiscale attention module EMA is introduced in the decoding stage to suppress the generation of noise and artifacts, potentially improving the generation effect. The experimental results show that on the multilingual offline handwritten character dataset, the CDE-DDPM model achieves a recognition accuracy of 93.03%. It is superior to other methods. The effectiveness of this method is also verified on other sample sets, and satisfactory results are achieved.</p>