Spoken Function Calling: A New Perspective on Spoken Language Understanding for Large Audio Language Models

This paper introduces Spoken Function Calling (SFC), a structured formulation of spoken language understanding intended to move task-oriented dialogue beyond closed-set intent and slot prediction. The authors extend spoken functions from established SLU datasets, use a multi-agent system to synthesize SFC-Bench, evaluate both LLMs and large audio language models, and apply post-training to improve SFC performance. According to the supplied abstract, SFC yields higher semantic extraction accuracy than traditional SLU formulations, although no benchmark sizes, model-level results, or numerical improvements are provided.