Hazard regression for modeling conversational silence

It is often assumed that the participants of a conversation try to avoid simultaneous starts or lengthy silences. For this reason, they may tend to synchronize rhythmically with each other’s speech. A model of conversational turn-taking based on the idea of coupled oscillators has been suggested by Wilson & Wilson [1]. However, the model has received only weak empirical support from previous studies where distributions of silence durations have been modeled directly. In the present study, we attempt to detect signs of oscillatory behavior during silence utilizing nonparametric hazard regression. In order to understand the shape of the estimated hazard rates, we postulate a latent stochastic process [2] with end of silence occurring when the process crosses a threshold. This finer-grained approach using Bayesian estimation yields a more detailed picture of synchronization between speakers and a more powerful test of oscillatory behavior.