Preliminary Results on Audio Event Classification Applied to a University Square in Milan (Italy) Before an Urban Regeneration Project

Biophonies, anthropophonies and geophonies characterize and shape an environment and contribute to the human appreciation of that place. Thus, sound event classification can be a useful tool to assess its quality and detect changes affecting it. In this study, different machinelearning models for multilabel sound classification are tested to monitor the ante-opera situation of the renewed square “Piazza della Scienza” of the University of Milano- Bicocca (Italy). The one-week monitoring was performed in May 2023 using 7 Song-Meter-Micros. The recordings were equalized to correct the devices’ nonlinear frequency response. The paper is structured to: (a) test two sets of features in the Piazza’s polluted soundscape by constant ventilation noise and other anthropogenic sources: YAMNet embeddings and classic audio features (such as MFCCs), (b) find the best algorithm between: decision tree, random forest, k-nearest neighbor and support vector classifier and (c) evaluate their performance when filtering the background ventilation noise to increase the datasets size. Preliminary results are presented with the final aim of optimizing the detection and applying it to describe the Piazza’s soundscape, investigate differences in events spatial distribution, and evaluate the effects of the urban regeneration plan on the soundscape.