Speaker-Independent Recognition of Negative Nonverbal Behaviors for AI-Based Oral Presentation Assessment

Presentation is a core competency required in organizational reporting, proposals, and interviews, yet presenters find it difficult to Precognize their own nonverbal behavior and to receive consistent feedback. This study develops and evaluates a model that recognizes negative nonverbal behaviors during presentations and classifies them into twelve categories, to support AI-based oral presentation feedback. Unlike existing systems that remain at holistic scores or the detection of a few behaviors, the model classifies a presenter's skeletal motion into individual behavioral units, providing a basis for specific feedback. The analysis uses nonverbal keypoint data from 800 presenters in AI-Hub's Public Speaking Practice and Evaluation Data and applies a speaker-independent evaluation in which no presenter's data appear in both training and evaluation. Deep-learning models achieved stable performance of around 90% macro-F1 on presenters unseen during training, and in this task the composition of the evaluation split affected the reported performance more than the choice of model did. A lightweight model remained within one percentage point of the best-performing model in accuracy, with classification latency low enough that this stage would not be the bottleneck for real-time feedback. These results provide a technical foundation for AI-based oral presentation feedback to serve as a fine-grained behavioral feedback tool in management education and organizational communication competency development.