Radio frequency (RF) signal recognition plays a critical role in modern wireless communication and security applications. Deep learning-based approaches have achieved strong performance but typically rely heavily on extensive training data and often fail to generalize to unseen signals. In this paper, we propose a method to learn discriminative embeddings without relying on real-world RF signal recordings by training on signals of synthetic wireless protocols. We validate the approach on a dataset of real RF signals and show that the learned embeddings capture features enabling accurate discrimination of previously unseen real-world signals, highlighting its potential for robust RF signal classification and anomaly detection.