Glimpsed Periodicity Features and Recursive Bayesian Estimation for modeling attentive voice tracking
* Presenting author
Computational models are a way of approaching research questions related to auditory perception. One relevant question is how we are able to follow and understand speech in complex acoustic scenes. Previous studies suggest that for tracking a speaker in such conditions, humans use (1) sparse, speaker-related bits of robust information - 'auditory glimpses' and (2) a mechanism of predictive coding with a movable locus of attention. The goal of the present study is to develop a computational model for attentive tracking of voices, which takes these two aspects into account. We model auditory glimpses using Glimpsed Periodicity Features, and predictive coding using Recursive Bayesian Estimation. We assume that perception is organized into an attended foreground and unattended background. We propose parallel particle filters - one for each category - to track the concurrent events. In this approach, each incoming glimpse is associated with either foreground or background based on accumulated evidence. Simulations with artificially generated data of a 'glimpsing' nature (sparse, robust) showed that this approach is suitable to track multidimensional parameter trajectories of two competing sources. This suggests the potential of the method to track simultaneously active voices based on the Glimpsed Periodicity Features.