Analysis of representation and generalization capabilities of pre-trained audio models in urban environments

In the last decade, urban noise pollution has become a significant environmental concern that can be mitigated with the help of audio detection algorithms for classifying different sources of noise and creating more informative noise maps. In this context, machine learning, particularly transfer learning, is an essential technology that enables accurate analysis of urban noise sources. However, the choice of the pre-trained model used to compute audio embeddings can significantly influence the performance of downstream classification tasks. This paper aims to compare the embeddings of various pre-trained models on different data collection campaigns in the context of the Sons al balcó project and quantify the robustness of audio representations. To achieve this, we develop metrics and statistically test the presence of distribution shifts in learned latent features. To evaluate the quality of the embeddings, we perform both qualitative and quantitative analysis using dimensionality reduction methods and assess the performance on downstream tasks using data from different collection campaigns. Results highlight major differences between general purpose and specific models. Our findings suggest the need for careful consideration during the choice of the pre-trained model to use in audio event detection applications.