NAACA: Training-Free Audio Model Boosts Salience Detection

Researchers have unveiled NAACA, a novel NeuroAuditory Attentive Cognitive Architecture designed to enhance the performance of audio language models in lengthy recordings by treating attention distribution as an auditory salience filtering challenge. Central to this architecture is OWM, a neuro-inspired Oscillatory Working Memory that sustains stable attractor-like states, activating higher cognitive processes only when adaptive energy variations indicate perceptual salience. On the XD-Violence dataset, NAACA boosted AudioQwen's average precision from 53.50% to 70.60%, simultaneously minimizing unnecessary ALM activations. Additionally, qualitative analyses of the Urban Soundscapes of the World (USoW) dataset showcased OWM's capacity to detect new events and subcategory transitions while remaining resilient to brief pauses.