Abstract: Detecting and identifying sound events is a core task with many applications in music. Supervised deep learning-based solutions are effective but require labeled datasets. Finding annotated music datasets is difficult for specific tasks at hand. Moreover, the performance degrades when models trained on synthetic datasets are deployed on real-world audio. We are working towards effective unsupervised or semi-supervised domain adaptation techniques for the above problem. Initial experiments show promising results. The domain adaptation methods we discuss here are fairly general and could be applied to other problems in music information retrieval too.