This paper proposes a centralized role-based architecture that utilizes Large Language Models (LLMs) to enable adaptive communication systems. The framework implements dynamic role-based LLM assignment, where robots receive specific language models based on their functional identity and communication objectives. A user study evaluating different LLM configurations demonstrates that role-specific LLM significantly enhances interaction clarity, contextual awareness, and perceived human likeness compared to uniform deployment. These findings establish that multi-model LLM integration improves collaborative social robots’ authenticity and functional efficiency in SHMR systems.