Software-defined radio (SDR) is a promising non-invasive approach for human activity recognition. While the deep learning methods in SDR-based HAR are of growing interest, the comparison of different model architectures has a lack of systematic empirical evidence describing the relative performance of different model architectures with the same signal conditions. Accordingly, this investigation performs an empirical evaluation of several deep learning architectures and classical machine learning architectures based on a publicly available SDR dataset. The publicly available University of Glasgow dataset, which comprises SDR devices and Universal Software Radio Peripheral (USRP) models X300/X310, was utilised to collect the data on the aforementioned activities and subsequently preprocessed and fed into a classifier. Five classifiers were systematically instantiated and evaluated: Convolutional Neural Network (CNN), one-dimensional Residual Network (1D ResNet), Long Short-term Memory (LSTM) network, Decision Tree and a Conditional Generative Adversarial Network (cGAN)-based classifier. Performance metrics were measured through overall classification accuracy since the preprocessing regimes and training regimes were consistent for all models. Experimental results show that the cGAN-based model achieved the highest accuracy of 96.4%, and CNN and Decision Tree show the close accuracy of 95.36% and 94.1%, respectively. Again, the performance of 1D ResNet was 86.2%, and that of LSTM was comparatively less at 75%. These results highlight the power of convolutional and adversarial models in learning discriminative signal features from the signal representations of the SDR, which, compared to purely sequential architectures, such as LSTM, demonstrate its limitation of the complex dynamics of radio frequency signals.