WoW-Bench: Evaluating Fine-Grained Acoustic Perception in Audio-Language Models via Marine Mammal Vocalizations
WoWBench introduces a new benchmark to rigorously test Large Audio Language Models' ability to perform low-level listening and cognitive processing using marine mammal vocalizations. The benchmark assesses perception and cognition through tasks ranging from categorizing sounds to analyzing complex acoustic sequences, revealing current models struggle with genuine perceptual grounding.