What makes audio event detection harder than classification?

Audio event classification and detection (AEC/D) have been an active field of research in recent years [1]–[3]. So far, beside a majority of works focusing on the improving overall performance in terms of accuracy [2], [1], [4], [5], many other aspects have also been studied, including noise robustness [6]–[7], [8], overlapping event handling [9], [10], [11], [12], early event detection [13], multi-channel fusion [14], as well as generic representation [15]. However, little attention has been paid to the important aspect of event detection systems on continuous streams: false positive reduction. False positives, i.e., event instances that are spuriously detected by a detection system, and subsequently draw attention to them, are arguably one of the most important problems faced by different applications like ambient intelligence and surveillance. To the best knowledge of the authors, this is the first work explicitly addressing this problem.