Frequency Tracking Features for Domain-Robust and Data-Efficient Deep Siren Identification

Siren identification systems are crucial for smart vehicles to early detect emergency situations that demand attention and, often, a response from the driver. The problem of recognizing the sound of a siren within traffic noise has been largely addressed using deep learning, to tackle the strong and unstructured background noise characterizing the urban soundscape. While convolutional neural networks processing spectrogram features are state-of-the-art, their performance is tied to the availability of large and diverse datasets and is hindered by domain-shift. In order to build domain-robust models that do not demand extensive training datasets, that are hardly available in practice, in this work we propose a low-complexity feature extraction method based on frequency tracking. Relying on the harmonic nature of siren signals, we adopt a single-parameter adaptive notch filter to track their periodically evolving fundamental frequency and the ratio between the power of the tracked sinusoid and that of the full input signal. These features are used as input for a small-scale convolutional network, trained entirely on synthetic data. Evaluation results indicate that the proposed system achieves a good generalization to multiple real-world datasets unseen during training, outperforming spectrogram-based state-of-the-art methods when limited data is available while also having a smaller model size.