Semi-supervised learning for automatic audio events annotation using TSVM

Most previous approaches to automatic audio events (AEs) annotation are based on supervised learning which relies on the availability of a labeled corpus to train classification models. However, instance annotation is often difficult, expensive, and time consuming. In this paper, we apply semi-supervised learning with transductive Support Vector Machine (TSVM) algorithm to automatic AEs annotation. Besides, considering about the presence of outliers which degrade the generalization and the classification performance, we propose a confidence-based method for samples selection. In our experiments based on the melodrama Friends corpus, the proposed method can effectively use unlabeled data to improve the classification performance with only a small amount of the labeled data.