Abstract : Recent breakthroughs in deep learning have contributed to this area of glyph classification, script recognition, and even multi-lingual Optical Character Recognition. Several deep learning architectures-from the conventional Convolutional Neural Network to the recent Vision Transformer and even their combination-have been proposed in the past for performing the mentioned tasks. Each of these deep learning architectures has varying strengths and weaknesses w.r.t. accuracy, complexity, scalability, and applicability. This paper presents a clear, comparative study of recent deep learning-based architectures on existing standard benchmarks like EMNIST, Omniglot, and multi-lingual script datasets. The study has brought forth the relevance of different architectures with respect to accuracy, complexity of calculations, execution speed, and practical applicability in applications related to a few-shot learning problem and edge computing.