Characterizing urban acoustic environments by clustering sensor node measurements

Equivalent sound pressure levels have been widely used to assess the impact of the sonic environment on persons. At the same time, it has been demonstrated that sound sensor networks allow to extract a wealth of information about the living environment. Nevertheless it has not been clearly demonstrated how to summarize this information. To this end, we used advanced non-linear clustering methods to group 15-minute sound environments based on a wide range of classical acoustic indicators such as peak levels, statistical levels, spectral center of gravity. Machine learning methods identifying the sound sources that are present based on 1/3 octave band levels measured at 125 msec interval, are then used to better understand why these clusters emerge. The methodology is applied to 150 bedroom window measurements gathered in the Equal-Life project. Preliminary results show that groups of pre-school children with specific cognitive development and mental health are related to different diurnal patterns of sound environment clusters. These results can be interpreted in the larger context of living environments extracted from environmental sound sensing networks.