In this paper, we model dyadic human conversational interactions from a nonlinear dynamical systems perspective. We focus on deriving measures of the underlying system complexity using the observed dyadic behavioral signals. Specifically, we analyze different measures of complexity in prosody of speech (pitch and energy) during dyadic conversations of couples with marital conflict. We evaluate the importance of these measures as features by correlating them with different behavioral attributes of the couple codified in terms of behavioral codes. Furthermore, we investigate the relation between the computed complexity and outcomes of couples therapy. The results show that the derived complexity measures are more correlated to session level behavioral codes, and to the marital therapy outcomes, compared to traditional speech prosody features. It shows that nonlinear dynamical analysis of speech acoustic features can be a useful tool for behavioral analysis.