Making sense of periodicity glimpses in a prediction-update-loop - a computational model of attentive voice tracking

Humans are able to follow a given speaker even in challenging acoustic conditions. The perceptual mechanisms underlying this ability remain unclear. In this study, we present a computational model of attentive voice tracking, consisting of four main computational blocks: A) sparse periodicity-based auditory feature extraction, B) foreground-background segregation, C) state estimation and D) top-down knowledge. Conceptually, the model brings together ideas related to auditory glimpses, foreground-background segregation and Bayesian inference. Algorithmically, it combines sparse periodicity features, sequential Monte Carlo sampling and probabilistic voice models. We evaluate the model by comparing it with the data obtained by listeners in the study of Woods and McDermott (2015), which measured the ability to track of one of two competing voices with time-varying parameters (fundamental frequency (F 0) and first two formants (F 1, F 2)). We simulate two experiments: Stream Segregation of Sources Varying in Just One Feature and Effect of Source Proximity. In both experiments, we test three model versions, which differ in the type of information used in the segregation stage: version 1 uses oracle F0, version 2 uses estimated F0 and version 3 uses spectral shape derived from estimated F0 and oracle F1 and F2. Version 1 simulates optimal human performance in the conditions with the largest separation between the voices, version 2 simulates conditions where the separation between the voices in not sufficient for humans to follow the voices, and version 3 is closest to human performance for moderate separation between the voices.