Semi-Supervised Sound Event Detection with Pre-Trained Model

Sound event detection (SED) is an interesting but challenging task due to the scarcity of data and diverse sound events in real life. In this paper, we focus on the semi-supervised SED task, and combine pre-trained model from other field to assist in improving the detection effect. Pre-trained models have been widely used in various tasks in the field of speech, such as automatic speech recognition, audio tagging, etc. If the training dataset is large and general enough, the embedding features extracted by the pre-trained model will cover the potential information in the original task. We use pre-trained model PANNs which is suitable for SED task and proposed two methods to fuse the features from PANNs and original model, respectively. In addition, we also propose a weight raised temporal contrastive loss to improve the model's switching speed at event boundaries and the smoothness within events. Experimental results show that using pre-trained model features outperforms the baseline by 8.5% and 9.1% in DESED public evaluation dataset in terms of polyphonic sound detection score (PSDS).