Adversarial Domain Adaptation for Open Set Acoustic Scene Classification

Many algorithms classify acoustic scenes with predefined acoustic scenes categories but few addresses identifying acoustic scenes that are not predefined (usually referred as “unknown acoustic scenes”), which is known as “open set” problem for acoustic scene classification. Traditional methods generally use a “one-size-fits-all” threshold to make a second judgment on the output of trained model. The boundary between known and unknown scenes cannot be learned. To enable this boundary to be programmed, this paper proposes a novel method to introduce adversarial domain adaptation into the open set acoustic scene classification. In this method, known scenes are classified through the adaptation of target domain and source domain, and unknown scenes are distinguished by adversarial training with the help of preset pseudo-threshold. Not only the discrimination between unknown classes and known classes can be learned during the adversarial training process, but the overall performance of the open set acoustic scene classification algorithm is also improved. The proposed system achieves better performance compared with the baseline of open set acoustic scene detection in Detection and Classification on Acoustic Scenes and Events challenge 2019.