Automatic Prediction of Disturbance Caused by Interfloor Sound Events

There is a direct correlation between noise and human health, while negative consequences may vary from sleep disruption and stress to hearing loss and reduced productivity. Despite its undeniable relevance, the underlying process governing the relationship between unpleasant sound events, and the annoyance they may cause has not been systematically studied yet. In this context, this work focuses on the disturbance caused by interfloor sound events, i.e., the audio signals transmitted through the floors of a building. Activities such as walking, running, using household appliances or other daily actions generate sounds that can be heard by those on an adjacent floor. To this end, we implemented a suitable dataset including diverse interfloor sound events annotated according to the perceived disturbance. Subsequently, we propose a framework able to quantify similarities exhibited by interfloor sound events starting from standardized time-frequency representations, which are processed by a Siamese neural network composed of a series of convolutional layers. Such similarities are then employed by a $k$-medoids regression scheme making disturbance predictions based on interfloor sound events with neighboring latent representations. After thorough experiments, we demonstrate the effectiveness of such a framework and its superiority over popular regression algorithms. Last but not least, the proposed solution offers interpretable predictions, which may be meaningfully utilized by human experts.