Humor Style Transfer for Virtual AI Teachers Using StyleGPT4o for Comedian Guided Script Rewriting in Virtual Educational Environments

This study proposes a virtual AI teacher humor style transfer model StyleGPT4o for virtual education environments, aiming to achieve high-fidelity transfer and voice output adaptation of the language style of specific characters in teaching scripts. Based on GPT-4o, the model automatically collects comedian corpus, constructs style vectors, and combines structured prompt control to guide the model to generate teaching texts with the humorous style of the characters. Without fine-tuning the original parameters, StyleGPT4o realizes the coordinated control of semantics and style, and supports downstream speech synthesis modules to convert text output into virtual speech with rhythm and emotional expression for voice-driven and interactive interpretation of virtual AI teachers. The experiment randomly selected 300 TED texts from more than 4,000 TED texts, constructed a parallel comparison task between StyleGPT4o and the original GPT-4o, and conducted a comprehensive evaluation from three dimensions: semantic fidelity, style transfer strength, and humor structure density. The results show that StyleGPT4o achieves stronger style consistency and output stability while maintaining the integrity of teaching semantics, and is suitable for personalized speech generation tasks of virtual AI teachers in virtual education scenarios.