Abstract
Gesture synthesis has gained significant attention as a critical researchfield, aiming to produce contextually appropriate and natural gesturescorresponding to speech or textual input. Although deep learning-basedapproaches have achieved remarkable progress, they often overlook the richsemantic information present in the text, leading to less expressive andmeaningful gestures. In this letter, we propose GesGPT, a novel approach togesture generation that leverages the semantic analysis capabilities of largelanguage models , such as ChatGPT. By capitalizing on the strengths of LLMs fortext analysis, we adopt a controlled approach to generate and integrateprofessional gestures and base gestures through a text parsing script,resulting in diverse and meaningful gestures. Firstly, our approach involvesthe development of prompt principles that transform gesture generation into anintention classification problem using ChatGPT. We also conduct furtheranalysis on emphasis words and semantic words to aid in gesture generation.Subsequently, we construct a specialized gesture lexicon with multiple semanticannotations, decoupling the synthesis of gestures into professional gesturesand base gestures. Finally, we merge the professional gestures with basegestures. Experimental results demonstrate that GesGPT effectively generatescontextually appropriate and expressive gestures.