Tension in communication often prevents effective flows of information among members in the conversation and thus can negatively influence practices in teams and learning efficiency at school. Interested in developing a computational technique that automatically detects tension in communication, we explore features that signal the existence of tension in human-human communications and investigate the potential of a supervised learning approach. While there is no tension-annotated dataset available, there are language resources that have distress annotated. Although tension may occur during the communication as a result of various factors, distress creates discomfort and tension. Leveraging an interview dataset that has marked the presence/absence of distress, we investigated the prosodic features and LIWC features that indicate tension. Specifically, we compare 23 prosodic features and LIWC features extracted from 186 interviews in terms of how effective they are to indicate the speaker’s distress in a one-to-one conversation. Our analysis shows that there are seven prosodic features and one LIWC features that differ between distress and non-distress interviews. The seven prosodic features are mean intensity, jitter, shimmer, longest silence duration, longest silence position, standard deviation of interviewee speaking rate, and hesitation. And the one effective LIWC feature is health.