Sign language is the primary means of communication for individuals with hearing impairments. Despite significant advances in assistive technologies, sign language remains widely used due to its efficiency, expressiveness, and natural integration into daily interactions. Developing automatic recognition systems is therefore essential to bridge the communication gap between sign language users and the broader hearing community. In this work, we focus on Arabic Sign Language (ArSL) recognition using the recently introduced KArSL database. Three deep learning approaches are investigated and compared: (i) spatiotemporal feature extraction using a 3D Convolutional Neural Network (C3D) followed by a multilayer perceptron (MLP) classifier, (ii) a 3D convolutional autoencoder designed for joint representation learning and classification, and (iii) a hybrid approach that combines the strengths of both methods by feeding the extracted C3D features into an autoencoder for dimensionality reduction and subsequent classification using a 1D-CNN. Experiments were conducted on a subset of 50 ArSL classes, enabling controlled evaluation and comparison between the proposed architectures. The obtained results demonstrate strong recognition performance, with each approach exhibiting distinct strengths in feature representation and classification capability. These findings confirm the effectiveness of deep learning-based architectures for Arabic Sign Language recognition and provide a solid foundation for further investigation on larger-scale datasets in future work.