Abstract
Labeling onset and offset time and class information for audio files has always been a laborious task for sound event detection (SED) research. Therefore, it would be appealing to consider using strongly-labeled synthetic data and a large amount of unlabeled data when training a SED model. To utilize these data, previous research showed that mean-teacher based semi-supervised consistency training strategies could work with adversarial domain adaptation. However, the effectiveness has only been demonstrated in the domestic environment. In this work, we expand the scope of application to bird sound event detection in the natural environment. To achieve this, we established a synthetic strongly-labeled bird sound dataset consisting of 20 species. Then, we tuned the consistency training strategies and domain adaptation techniques that have been proven effective in the domestic environment to this new dataset. Our evaluation on the Eastern North American bird sound dataset resulted in an F1 score of >30%, which approaches the upper-bound performance that could be obtained when strongly-labeled data are assumed to be available. Thus, the strategies exhibit potential for advancing research in both indoor and outdoor SED.