This study aims to develop a handwritten English character recognition model by training a compact multilayer perceptron (MLP) neural network with text strokes. The primary data for model training was stroke features, including stroke orientation and density. Feature extraction was used to achieve sampling and data reduction. Training time was significantly reduced by using data reduction, where each iteration ran about 10 times faster than that using LeNet-5. With less training time, the model can make faster predictions. This model is suitable for edge computing and real-time prediction, even for Internet of Things (IoT) platforms not supported by graphics processing unit (GPU). This study used one split of the extended MNIST (EMNIST) dataset, i.e., EMNIST-Balanced, which consists of 47 classes containing English digits. According to literature related to EMNIST, the classification accuracy with this dataset was 78.02%. In comparison, this study achieved an accuracy of 87.49%