Semi-Supervised learning for multi-label audio tagging

Abstract

Audio Tagging is concerned with the development of systems that are able to recognize sound events. A growing interest is geared towards audio tagging for various applications such as acoustic surveillance, tagging video content and environmental scene recognition. Our goal is to design an audio tagging system capable of recognizing a wide range of sound events. The development process usually requires a large set of labeled sound data. However, most existing datasets are unlabeled since hand-labeling is a very costly and a time-consuming process, and it involves a lot of manual labor. To mend with this, we have built our audio tagging system following the Semi-Supervised Learning (SSL) paradigm. Specifically, we have chosen the pseudo-labeling strategy to learn from weakly labeled data. In addition, our system trains a ResNet deep learning model on log-mel spectrograms, along with augmentation techniques to increase the dataset size. The training uses the cyclic cosine annealing technique for the learning rate. We have carried out our experiments on a huge dataset made of sound recordings; we have investigated the impact of the sharpening temperature (a hyperparameter of our system) on the distribution of the pseudo-labels, and have tested ensembling various variants of our approach. The results demonstrate the efficacy of pseudo-labeling SSL strategy. Furthermore, ensembling various systems significantly boosts the overall performance. Keywords: Audio Tagging, Semi-Supervised Learning, Feature Extraction, Deep Learning, Ensemble Learning, Statistical Tests.