Soundscapes on edge - The real-time machine learning approach for measuring Soundscapes on resource-constrained devices

According to the WHO, noise is a growing problem in urban areas and the second most common environmental cause of health issues in Europe. As a complementary approach to noise maps based on sound pressure levels, soundscape maps can be a useful tool for urban planning, providing more information about how people perceive acoustic environments. This study describes an in-situ soundscape monitoring system based on WASN (Wireless Acoustic Sensor Network) for statistical spatial-temporal prediction of soundscapes in urban areas. Soundscape data on specifically defined spatial scales were observed and evaluated using a microcontroller with a 32-bit nRF52840 Nordic Semiconductors CPU and 1MB of memory in a multifunctional urban area. The use of TinyML enabled machine learning algorithms provided state-of-the-art soundscape classification to a low-cost edge device with extreme resource constraints regarding memory, speed, and lack of GPU support. Sound source types are classified into anthrophony, traffic, biophony, and geophony sounds using ESC-50 for evaluation. Our final MFCC-based CNN achieved an accuracy of 81.6% and even reached higher accuracy in the enclosed studio test. The results show that it is computationally feasible to classify soundscapes on low-power microcontrollers and potentially inform decision-makers based on extended sound analysi