Audio event detection in the era of Big Data has the constraint of lacking annotations to train robust models that match the scale of class diversity. This is mainly due to the expensive and time-consuming process of manually annotating sound events in isolation or as segments within audio recordings. In this paper, we propose an approach for semi-supervised self-training of audio event detectors using unlabeled web data. We started with a small annotated dataset and trained sound events detectors. Then, we crawl and collect thousands of web videos and extract their soundtrack. The segmented soundtracks are run by the detectors and different selection techniques were used to determine whether a segment should be used for self-training the detectors. The original detectors were compared to the self-trained detectors and the results showed a performance improvement by the latter when evaluated on the annotated test set.
View on arXiv