FairTalk: Facilitating Balanced Participation in Video Conferencing by Implicit Visualization of Predicted Turn-Grabbing Intention

TL;DR

FairTalk addresses unequal speaking opportunities in video conferences by predicting turn-grabbing intentions using a positive-unlabeled learning framework and subtly conveying these intentions through a pseudo-leaning-forward visualization. The approach relies on a lightweight time-series model trained with automatically labeled positives derived from turn-taking events, avoiding heavy manual annotation. A controlled user study suggests implicit visualization can modestly improve speaking balance while subjective perceptions show no strong impact, indicating the technique may subtly influence dynamics without imposing cognitive load. The work offers design guidance for mindless, implicit feedback in virtual meetings and provides a public dataset to support reproducibility and future improvements.

Abstract

Creating fair opportunities for all participants to contribute is a notable challenge in video conferencing. This paper introduces FairTalk, a system that facilitates the subconscious redistribution of speaking opportunities. FairTalk predicts participants' turn-grabbing intentions using a machine learning model trained on web-collected videoconference data with positive-unlabeled learning, where turn-taking detection provides automatic positive labels. To subtly balance speaking turns, the system visualizes predicted intentions by mimicking natural human behaviors associated with the desire to speak. A user study suggests that FairTalk may help improve speaking balance, though subjective feedback indicates no significant perceived impact. We also discuss design implications derived from participant interviews.

Figures (6)

• Figure 1: We designed a system, FairTalk, that automatically facilitates video conferencing to promote a fair talking opportunity distribution. It predicts listeners' intention of turn-grabbing (willingness to speak) by a machine learning model trained on a novel unsupervised framework based on positive-unlabeled (PU) learning. An implicit visualization with pseudo-action helps participants unconsciously balance the talking opportunity. (Photos from https://vimeo.com/654390624)
• Figure 2: The flowchart of data processing.
• Figure 3: The distribution of speaking turns (left) and speaking time (right). Participants are ordered within each group from the shortest to the longest speaker for each metric.
• Figure 4: The distribution of responses for each Likert question item (1: strongly agree; 7: strongly disagree).
• Figure 5: An example of face feature sequence sampling.