QG-DETR: Query-Guided Detection Transformer for Audio Moment Retrieval

Audio Moment Retrieval (AMR) has garnered increasing attention due to the rising demand for efficient audio content understanding. AMR aims to accurately locate temporal intervals in untrimmed audio based on natural language queries. However, existing approaches often struggle with cross-modal alignment, primarily because they fuse textual queries with audio representations containing irrelevant content, thereby impairing retrieval accuracy. To address this, we introduce Query-Guided DETR (QG-DETR), a novel AMR framework incorporating two critical components: a Query-Guided Audio Enhancement Module, which leverages textual queries to refine audio representations by suppressing irrelevant information, and a Cross-modal Alignment Module, explicitly designed to fuse refined audio features and textual representations to ensure accurate localization of relevant audio moments. Extensive experiments conducted on benchmark datasets including Clotho-Moment, UnAV-100, and TUT Sound Events 2017 demonstrate that our QG-DETR achieves state-of-the-art performance.