The capability to discriminate signals in the wireless spectrum is significantly crucial in a wide range of practices, such as mesh networks, military, and defense applications. Motivated by these applications, we propose a signal classification scheme deployed in embedded software-defined radio. The proposed work is differentiated by designing the system for over-the-air signals and considering real-life conditions: hardware-constraints and varying channel congestions. The modified Bartlett-based method is unified with a lightweight CNN on a standalone embedded device. Datasets are produced by over-the-air measurements from a WLAN modem that reflects the sparse and dense channel scenarios. The influence of bit resolution and dimension count of the dataset on classification is monitored. We demonstrate the designed system performance across a variety of traffic rates to examine stability. Additionally, several state-of-the-art CNNs are compared with the proposed CNN for accuracy, model complexity, and FLOPS. Our CNN architecture demonstrates immensely nearby accuracy (0.7% - 2.1%) to the rivals despite its exceedingly lightweight architecture.