Stroke is a leading cause of gait disability that leads to a loss of independence and overall quality of life. The field of clinical biomechanics aims to study how best to provide rehabilitation given an individual's impairments. However, there remains a disconnect between assessment tools used in biomechanical analysis and in clinics. In particular, 3-dimensional ground reaction forces (3D GRFs) are used to quantify key gait characteristics, but require lab-based equipment, such as force plates. Recent efforts have shown that wearable sensors, such as pressure insoles, can estimate GRFs in real-world environments. However, there is limited understanding of how these methods perform in people post-stroke, where gait is highly heterogeneous. Here, we evaluate three subject-specific machine learning approaches to estimate 3D GRFs with pressure insoles in people post-stroke across varying speeds. We find that a Convolutional Neural Network-based approach achieves the lowest estimation errors of 0.75 ± 0.24, 1.13 ± 0.54, and 4.79 ± 3.04 % bodyweight for the medio-lateral, antero-posterior, and vertical GRF components, respectively. Estimated force components were additionally strongly correlated with the ground truth measurements ( ). Finally, we show high estimation accuracy for three clinically relevant point metrics on the paretic limb. These results suggest the potential for an individualized machine learning approach to translate to real-world clinical applications.