Features analysis to label sound sources by long-term monitoring

Abstract

Background noise, speech privacy, and productivity are strictly related in offices. Employees’ speech may affect the performance of colleagues disturbing their concentration. On the other hand, HVAC noise can mask irrelevant speech noise, improving in some way employee’s comfort. Thus, it becomes important to achieve the ability to individuate in an unattended way the different sound sources during working hours: mechanical (HVAC, devices, computers, …), human, traffic, activities in nearby spaces. Clustering techniques provide tools to find patterns among unlabeled data. In a previous study, Gaussian Mixture Model and K-means clustering were applied to sound level meter measurements carried out during working hours. The reliability of results encourages the investigation to find robust features to label the sound sources. In the present work, a sound level measurement has been carried out alongside the audio recording of the working activities throughout a whole day. Besides the clustering performed via Gaussian Mixture Model and K-means clustering over the sound pressure levels strings, a Variational Autoencoder was used to find latent features from the recording. Correlations between the methods explore the chance to obtain a broader understanding of data obtained by long-term monitoring.