Temporal Audio Grounding

Temporal audio grounding studies how models identify the time intervals that support descriptions, answers, or other claims about a recording. Approaches include frame- and event-level supervision, timestamp tokens and time-aware audio-language alignment, plus reinforcement learning with interval-based rewards; synthetic and pseudo-labeled data help scale training. Benchmarks cover dense sound-event captioning, speech and music, and long recordings where relevant moments must be found among overlapping or redundant audio. Models can recognize an event yet miss its timing, and precise localization remains difficult for brief events and long-form audio; gains from grounding training do not always transfer to broader understanding.