Unsupervised Inference of Auditory Attention from Biosensors

We study ways of automatically inferring the level of attention a user is paying to auditory content, with applications for example in automatic podcast highlighting and auto-pause, as well as in a selection mechanism in auditory interfaces. In particular, we demonstrate how the level of attention can be inferred in an unsupervised fashion, without requiring any labeled training data. The approach is based on measuring the (generalized) correlation or synchrony between the auditory content and physiological signals reflecting the state of the user. We hypothesize that the synchrony is higher when the user is paying attention to the content, and show empirically that the level of attention can indeed be inferred based on the correlation. In particular, we demonstrate that the novel method of time-varying Bayesian canonical correlation analysis gives unsupervised prediction accuracy comparable to having trained a supervised Gaussian process regression with labeled training data recorded from other users.