An Audio Attention Computational Model Based on Spatial Cues Gradient

Present bottom-up audio attention computational model extracts the underlying characteristics of single channel audio such as energy, pitch, zero crossing rate etc., to calculate the audio signal attention level. In audio surveillance, sound source whose directions change rapidly should have higher attention level. But present audio computational models cannot effectively express the audio attention caused by such signals. It may cause some audio events, which should be paid attention to, cannot bedetected. To solve this problem, based on the psychological principles that spatial information affects attention, this paper proposes a model that introducing the short-term spatial gradient cues to measure the attention caused by the fast changing of single audio source space direction. This model calculates the mean short-term changes of the spatial cues vector ofsub-bands as spatialcues gradient. Compared to the traditional audio attention computational model, the recall of detection of attention audio events increased 4.5 percentage points in experiments.