An AI-IoT Platform for Psycho-Acoustic Annoyance Assessment on the Edge

The Internet of Things (IoT) has revolutionized data gathering and analysis, providing valuable insights into various aspects of daily life. Noise pollution assessment, especially in urban areas, is crucial due to population growth and disturbing sounds. Psycho-acoustic parameters are essential for evaluating sound discomfort. Integrating these parameters with wireless acoustic sensor networks (WASNs) enables the creation of acoustic discomfort maps in smart cities. However, real-time monitoring of these parameters requires significant computational power, posing challenges for efficient node computations. To bypass this problem, it has traditionally been considered to send audio streams outside the WASN network for further calculation, generating traffic and security problems or the use of less accurate indicators that have a lower computational cost. To address these challenges, our work presents a solution based on a deep convolutional neural network (CNN). This CNN is trained on a dataset of typical sounds commonly encountered in urban environments. After training, the CNN can predict psychoacoustic parameters, defined by Zwicker’s psycho-acoustic annoyance model, with remarkable accuracy directly from the raw recorded audio signal. This enables real-time operation and continuous monitoring of the psycho-acoustic parameters using different IoT devices. Taking advantage of this fact, a field test has been performed in a real environment to verify the proper behavior of the IoT system, with excellent results and allowing a fast, simple and very intuitive acoustic monitoring.