Abstract: We consider the problem of using speech processing to characterize an aggregate of voice data, in contrast to inferences about individual voice cuts. We derive simple turn-taking models from speaker activity detection output on the Switchboard-1 corpus. These can be used to cluster speakers into turn-taking 'styles.' Demographic fields and turn-taking behavior prove to be statistically dependent, thus observed speaker activity improves estimates of the demographics of held-out data. Finally, we use turn-taking style to estimate speaker influence.