This study addresses the phenomenon where, despite the rapid advancement of Generative AI and Large Language Model (LLM) technologies contributing to the efficiency of resource production in the game industry, innovation in the actual player interacti...
This study addresses the phenomenon where, despite the rapid advancement of Generative AI and Large Language Model (LLM) technologies contributing to the efficiency of resource production in the game industry, innovation in the actual player interaction experience (UX) remains insufficient. In particular, existing voice recognition games rely on 'Keyword Spotting' methods that recognize only pre-defined specific commands. Consequently, they fail to accommodate players' free utterances and force the memorization of precise commands, thereby causing high cognitive load and hindering immersion. This study aims to overcome these structural limitations.
Accordingly, this study proposes a 'Generative AI-based Intent Inference Interaction System' that utilizes the advanced context inference capabilities of Large Language Models to interpret players' imperfect natural language utterances as 'Intent' rather than simple errors and connect them to specific ingame actions, verifying its effectiveness empirically.
To conduct this research, the rigid interaction problems of existing systems and the diverse utterance desires of players were first identified through literature review and Video Analysis. Based on this, a real-time testbed was established by integrating the Gemini API and Azure Speech API within an Unreal Engine 5 environment. Furthermore, to suppress AI Hallucination and output structured data processable by the game engine, a 'Goal-Oriented Constraint Prompt,' combining system goal definition and constraint setting, was designed through prompt engineering.
The results of a Usability Test conducted with 10 adults to verify the efficacy of the proposed system are as follows. First, quantitative data analysis showed that the proposed system improved task performance efficiency by approximately 51.8% compared to the existing keyword method, and drastically reduced the average number of attempts per unit task from 2.43 (SD=1.58) to 1.17 (SD=0.38) (t=8.006, p < 0.001). In particular, the significant reduction in standard deviation implies that the proposed system provides consistent interaction performance regardless of individual user differences or speech habits. Second, qualitative interview and speech data analysis confirmed that players' utterance patterns were not chaotic noise but could be systematized into three major categories: Pleonasm Speech, Deconstruction Speech, and Distort Speech, along with seven detailed sub-types. Participants experienced 'immersive chanting' as if they were actual wizards, freed from the compulsion to input correct commands. They showed a positive attitude change, perceiving the interaction itself as playful content, such as utilizing memes or intentionally distorting pronunciation.
In conclusion, this study demonstrates the feasibility of an 'Intent-Centric' interface that accepts players' natural language habits as they are. Based on this, the study presents a classification system for user utterance types, along with prompt engineering and UX design guidelines for applying Generative AI to games. The results of this study are expected to make academic and industrial contributions as a practical Reference Model for maximizing user experience beyond simple technical implementation in the future development of AI-based game content