Human vs. Computer performance in voice-based recognition of interpersonal stance

Abstract

This paper presents an algorithm to automatically detect interpersonal stance in vocal signals. The focus is on two stances (referred to as ‘Dominant’ and ‘Empathic’) that play a crucial role in aggression de-escalation. To develop the algorithm, first a database was created with more than 1000 samples from 8 speakers from different countries. In addition to creating the algorithm, a detailed analysis of the samples was performed, in an attempt to relate interpersonal stance to emotional state. Finally, by means of an experiment via Mechanical Turk, the performance of the algorithm was compared with the performance of human beings. The resulting algorithm provides a useful basis to develop computer-based support for interpersonal skills training.