Evaluating the Feasibility of Video Language Models in Assessing Nonverbal Cues

Presentational skills are learned primarily through constant feedback and practice. Without accessibility to experienced help, it can be difficult to spot deficiencies in the various facets of oral presentations, ranging from body language to intonation, to stance and movement. This paper investigates the feasibility of using emerging vision language models (VLMs) to evaluate nonverbal communication cues in oral presentations. We compared the performance of the top open and closed-source VLMs against the popularity of the presentations and we found that the models achieved a significant accuracy in following and regressing the like-to-view ratio of the videos. Fine-tuning the models, utilizing solely public speaking competition videos, and integrating verbal cues could further improve the accuracy and potential of the technology. These models can be applied to the world of competitive speaking environments; using them as an auxiliary tool can balance out possible subjective biases in judges that can complicate fair assessment.