In the modern privacy-sensitive environment, identifying a particular radio device among a larger pool of Internet-of-Thing (IoT) facilities has become more difficult than ever. In particular, the identification process is susceptible to security, mobility, and environment changes. Machine learning techniques, on the other hand, offer a unique learning mechanism to reconstruct corrupted signals from noise and interference at wireless transceiver chains, providing a stable and accurate system for applications on wireless communication. In this paper, we exploit a Deep Learning (DL)-based identification strategy for Radio Frequency (RF) fingerprints. Specifically, we introduce an Echo State Network (ESN), uniquely suited for nonlinear information processing, to discover dissimilar RF fingerprints directly from raw transmitted RF signals, allowing the network to discriminate radio devices. Through an over-the-air WiFi transmission dataset, numerical evalu- ations demonstrate advantages of the ESN over state-of-the-art DL-based identification approaches, yielding an average classification accuracy of 98.11% while significantly reducing the training overhead. Furthermore, compared to our baseline models with various DL architectures, the ESN can be trained even with a very limited training set without degrading its performance.