AI VTuber Development Base on GPT-4 for Casual Chat Interaction on YouTube Live Streaming: Case Study KAIRA Channel

Abstrak

This study presents the development and evaluation of KAIRA, an Indonesian AI-based VTuber designed to engage in casual conversations during live streaming on YouTube. KAIRA integrates GPT-4 for language generation, ElevenLabs for voice synthesis, and VTube Studio for real-time avatar animation. The system was evaluated through both controlled and public testing sessions, focusing on interaction quality, contextual relevance, and character liveliness. Semantic similarity was analyzed using IndoBERT and IndoRoBERTa to assess the alignment between the system's responses and user expectations. Additionally, a topic filtering mechanism based on cosine similarity is implemented to ensure KAIRA remains focused on a topic, with the testing topic being gaming. Evaluation of this classifier using a manually labeled dataset yielded an accuracy of 50.33%, revealing significant classification errors. However, due to the system’s flexible routing logic, misclassified messages often still received contextually appropriate responses. These findings highlight the dual challenges of topic enforcement and conversational coherence in real-time AI systems and contribute to the growing field of virtual character development and conversational AI for Indonesian-language contexts.

Identifier

• OpenAlex: https://openalex.org/W4416055919
• DOI: https://doi.org/10.23960/pepadun.v6i2.285