Towards automated assessment of communication skills in virtual reality: an exploratory feasibility study

Virtual reality (VR) enables standardized yet socially realistic assessments by embedding participants in controlled interpersonal situations while capturing rich process data. We present an immersive stealth-assessment system for professional communication competence that integrates scenario design, multimodal sensing, and exploratory automated scoring. Ninety-six adults with at least 2 years of professional experience completed a Unity-based virtual workplace module delivered on a Meta Quest Pro head-mounted display with integrated eye tracking. Because each automated-scoring outcome required a valid criterion label and a complete usable set of corresponding multimodal predictors, the outcome-specific analytic samples varied from N = 92 to N = 95. Across three connected scenarios—a project meeting, a one-to-one debriefing, and a high-stakes executive presentation—participants completed scripted interactions with 3D virtual characters and a 360° panoramic presentation context while gaze, speech audio, paraverbal descriptors, automatic transcripts, and event logs were recorded. In this revision, the automated-scoring analyses are explicitly framed as internal feasibility evidence rather than as validation of an operational assessment instrument. We report the Low/Medium/High cut-offs and class distributions, compare accuracy with empirical majority-class baselines rather than relying only on a balanced 0.33 benchmark, and identify outcomes with very small categories. Because the retained dataset does not allow a defensible retrospective estimate of inter-rater reliability for behaviorally rated outcomes, those outcomes are interpreted as exploratory labels and not as reliability-supported criterion measures. Selected-model accuracies were descriptively above empirical majority baselines for several outcomes, but accuracy alone is insufficient to establish model quality under class imbalance. The study therefore demonstrates the feasibility of multimodal evidence generation in VR while specifying the reliability, class-sensitive performance, external validation, and fairness analyses required before any high-stakes or operational use.