SAT-SED: Semi-Supervised Sound Event Detection by Pseudo-Labeling with Self-Adaptive Threshold Strategy

Semi-supervised sound event detection (SED) methods based on consistency regularization and pseudo-labeling have achieved significant success by leveraging unlabeled data for training. However, they still exists the issues of incorrect pseudo-label and error accumulation, resulting that the model may train along the wrong direction. In this paper, we propose a self-adaptive threshold training strategy for semi-supervised SED (SAT -SED), it adjusts the confidence threshold in a self-adaptive manner on the basis of the learning status of the model. The SAT strategy is composed of self-adaptive clip- (SACT) and frame-wise threshold (SAFT) strategies. As the SAT strategy is exclusively employed during the training phase, it improves the system's performance without increasing the model's parameters and inference time. We conducted the experiments on the development set DESED of DCASE2022 task4. The results of the experiments demonstrate that SAT-SED achieves considerable performance. Furthermore, we extended the SAT to a powerful algorithm framework, whose performance is further increased.