SPICA: Scalable and Personalized Conversational Agent Framework for AAC Users

Augmentative and Alternative Communication (AAC) users face persistent challenges in expressing themselves authentically. The effort required to compose messages and sustain conversational flow often prevents users from fully participating in natural dialogue. Previous works have explored the integration of large language models to reduce effort and accelerate communication. However, these systems often fail to capture the user's personal voice. To address this, researchers have explored fine-tuning with user data, yet these methods remain difficult to scale and generalize poorly beyond biographical content. In this work, we introduce SPICA, a unified framework that addresses two key limitations: (1) the lack of scalable personalization that can adapt to user contexts in real time, and (2) the absence of agentic mechanisms to organize and orchestrate knowledge for conversation. SPICA acts as a lightweight plug-in that dynamically indexes and restructures user-relevant information into a personalized knowledge base. Beyond indexing, SPICA retrieves relevant knowledge on demand to guide conversation. It enables responses that are faithful to the user's identity while remaining flexible for broader communication goals. We validated SPICA extensively through automated evaluations using 200 synthetically generated AAC user profiles, as well as qualitative studies with AAC users in real-world settings. Results demonstrate that SPICA enables faster communication while preserving personalization, producing responses that are contextually grounded and aligned with each user's unique style.