VARIATIONAL AUTOENCODER AND CLUSTERING TECHNIQUES TO SEGREGATE SOUND SOURCES

The separation and analysis of sound sources play a key role in assessing noise exposure in work environments. Sound level meter analyses do not allow technicians to quantify the noise's contribution of each sound source in a manifold context. This reveals a great lack of detail in plenty of cases. In office spaces, for example, different noise components can increase or decrease the employees' productivity. Machine and deep learning provide techniques to analyze large volumes of data. A previous study proposed two clustering techniques, the Gaussian Mixture Model and K-means, to analyze long-term measurements of the working activity within offices. In the present work, the daily activity of an office is analyzed through unsupervised clustering techniques and a semi-supervised analysis via a Variational autoencoder model. The first method exploits the sound pressure levels obtained by a sound level meter. The second method exploits the spectrograms obtained by the digital audio recording of the same activity, to obtain more detailed information about the sources. Results show good performance of both methods to identify and separate the traffic noise from the speech signal. The latent space of the Variational autoencoder can be a qualitative validation tool for the clustering methods.