Humanoid robots equipped with advanced natural language processing capabilities hold immense potential for various applications, yet face significant challenges in achieving human-like communication due to limitations in current Large Language Model (LLM) implementations. To address this, we propose a novel hybrid approach that integrates multiple LLMs with edge computing and caching mechanisms, consisting of four key steps: query analysis, model selection, response generation, and caching. Our main contributions include a novel edge computing architecture that optimizes human-robot interaction processing, an intelligent caching system for efficient query handling, and a dynamic model selection framework that adaptively routes queries to appropriate LLMs, collectively enabling faster and more natural robot-human communication. Experimental results exhibit the superiority of the proposed method in comparison with competitive baselines, in terms of both quantitative and qualitative manners.