Identification of Anomalies in Urban Sound Data with Autoencoders.

The growing population in the metropolises is influencing the need to plan cities to be safer for people. Several Smart Cities initiatives are being implemented in the cities to achieve this goal. A network of acoustic sensors has been deployed in New York City thanks to the SONYC project. Sounds of the city are being collected and analyzed. In this research work, acoustic signal data are represented with Mel-spectrogram images with mel-scale frequency versus time on a decibel scale. Traditional autoencoders and variational autoencoder models are deployed to detect anomalies in the mel-spectrogram images. The obtained results demonstrate that the variational autoencoder model finds anomalies accurately in the acoustic records.