2026
EMNLP
EMNLP 2026
Can Audio LLMs Understand Prosodic Focus? An Alternative-Semantics Inference Test with "Only"
Abstract
AbstractWe introduce a new inference test for audio LLMs, where the correct response crucially depends on the location of a focal accent. The proposed task shows for the first time how to test the ability of LLMs to incorporate audio information in truth-conditional semantic interpretation, relying on the theory of alternative semantics. Several native audio models are tested, with performance ranging from random guessing to around 90% accuracy. The task is challenging for all the models tested. A detailed error analysis reveals specific areas of difficulty. The theory of alternative semantics provides a basis for a wide range of systematic tests of audio models.