Reviews: Latent Gaussian Activity Propagation: Using Smoothness and Structure to Separate and Localize Sounds in Large Noisy Environments

This paper describes a system for localization and separation of sound sources recorded by distributed microphone arrays. It is similar to MESSL (Mandel et al, 2009) in its building blocks, but builds upon it by employing a Gaussian process prior to encourage smoothness in time, frequency, and location. It also is targeted to the same application as DALAS (Dorfan et al, 2015) of microphones being distributed in pairs around a large space, but adds the smoothness over time and frequency via the Gaussian process prior. It is evaluated on synthetic sound mixtures that are anechoic, but contain a reasonable amount of additive uncorrelated noise. Its performance is quantified using a new metric that compares the masks it estimates at each microphone to the ideal ratio mask at each microphone.