Smart noise monitoring at live concerts: Combining deep learning and acoustic sensor networks

Exposure to high sound levels at leisure setting such as live music events is a recognized public health concern and is subject to increasingly strict regulatory limits. Ensuring compliance with these limits requires continuous noise monitoring, often performed using networks of connected acoustic sensors that provide real-time measurements during events. A challenge in this context is that sound level exceedances may be caused not only by the concert itself but also by unrelated external noise sources, making manual verification by technicians necessary and limiting the scalability of current monitoring practices. This paper investigates the use of automated acoustic event classification to distinguish noise generated by live music performances from external and non-event-related sounds in complex, real-world acoustic environments. We present the collection and human annotation of acoustic data recorded during three large-scale music festivals and evaluate the use of Pre-trained Audio Neural Networks (PANNs) for this task. Several model configurations and backbones are explored, including the use of PANNs as feature extractors and as fine-tuned classifiers, with additional data augmentation to improve robustness against diverse background noises. Experimental results demonstrate that fine-tuned PANN-based models can reliably discriminate concert-related sounds from external noise, achieving F1-score and AUC values above 80% across different festival scenarios. These findings highlight the feasibility of integrating acoustic event detection into real-time noise monitoring workflows, reducing the need for manual validation and supporting more scalable and reliable regulatory noise assessment leisure settings.