YOLO-based Signal Detection of Amplitude Modulated Audio Transmissions in Realistic HF Scenarios

In this paper, we investigate a Deep Learning based approach to the detection of analog audio transmissions in the radio frequency (RF) spectrum with a particular focus on the noisy high frequency (HF) band. To this end, we adapt a self-trained YOLO-based detector for this demanding application scenario. The YOLO-based detector is subsequently compared to a previously proposed speech detector, designed for wideband RF signals, that is based on classical signal processing techniques. In our evaluation, we systematically evaluate both systems on synthetically generated signal scenarios. We then investigate how the suggested detector generalizes to real-world signals and examine the effect of fine-tuning on real-world data after an initial pre-training on synthetic data. Our paper highlights the benefits of the YOLO-based approach as compared to classical methods and contributes insights on the impact of training with synthetic versus recorded data for an application in real HF scenarios.