The Influence of the Interlocutor's Gender on the Speaker's Role Identification

The objective of the current on-going research is to automatically identify the role played by a speaker in a dialogue, and to explore potential conditions that might impose higher speaker's role identification. We use an interactive Map Task setup with two potential roles: followers and leaders, where each speaker participated twice thus acting in both roles with the same interlocutor. The paper aims to identify speaker's role, and to explore potential influence of the gender of the speaker, the gender of the interlocutor, and the order of the roles played by the speaker. By using deep learning procedures over a set of acoustic features, we automatically trace the footprints of the role through the speech signal. Results show an average of 73.3% role's classification rate. We further show that there is a significant difference in the role's classification rates, depending on the interlocutor's gender. On average, when the interlocutor is a male, the speaker tends to identify with his or her role more clearly - 77.5% versus 69.9% when the interlocutor is a woman.