In this paper, we propose generating synthetic time-series data to improve gesture classification using TabDDPM with a post-process filtering algorithm. We address the challenge of producing quality synthetic data for multivariate time-series datasets, where it is essential to capture spatial and temporal correlations for subsequent machine learning tasks. Both diversity and generalisation are limited by traditional oversampling methods like SMOTE and Random Sampling, which either introduce redundant samples or lack temporal awareness. In order to get beyond these restrictions, we suggest a pipeline that extracts unique temporal, statistical, and geometric properties from time-series data. We utilised and modified a diffusion-based generative model, TabDDPM. We present a post-processing approach that combines prototype filtering, classifier confidence, and Mahalanobis distance to choose high-quality synthetic samples to further improve label fidelity and sample realism. The Gesture Phase Segmentation dataset is used to test our method with five classes, and the results show that TabDDPM generated samples perform better than SMOTE and Random Sampling in terms of sample novelty (higher Mahalanobis Distance of 61.31, and 13.74 Distance to Closest Record). Furthermore, TabDDPM achieved 66% overall accuracy, surpassing the baseline and outperforming both SMOTE (64%) and Random Sampling (58%). Future research will focus on modifying the post-processing approach for direct application to data selection and integration with traditional augmentation methods.