Abstract
Conventional urban noise monitoring relies on aggregated A-weighted sound pressure levels, which provide limited insights into the present sound sources and temporal structure of the sound environment. This study presents results from a city-wide deployment of 38 low-cost, source-aware sound sensors in Delft, the Netherlands, accompanied by a survey on perceived sound prevalence and annoyance. The sensors measured loudness levels, intermittency ratios (IR), and the presence of 11 sound event classes, as predicted by an on-device machine learning model.The results reveal substantial spatial and temporal heterogeneity in the urban sound environment. Median loudness and intermittency varied strongly across locations, indicating that they represent complementary acoustic dimensions. Sound event presence ratios and causes of intermittency often differed from each other, highlighting the importance of distinguishing between overall source presence and their presence in the form of disruptive, intermittent events. Correlation analyses show that louder and event-like detections of vehicles and sirens are more strongly associated with reported annoyance than overall presence ratios. While source-specific intermittent detections relate to annoyance, the global IR did not correlate with perceived short-burst annoyance, suggesting that aggregated intermittency metrics may not directly translate into subjective experience.The findings demonstrate that source-aware, edge-based sensor networks provide insights beyond conventional dB(A)-based monitoring and offer a scalable and privacy-friendly approach for urban sound monitoring. The deployment and analysis provided valuable details about Delft’s soundscape. Future research should validate perceptual associations in larger samples and evaluate potential inaccuracies in machine learning-based sound event detections.

This publication has 20 references indexed in Scilit: