Dictionary-based active learning for sound event classification

This paper proposes a new dictionary-based active learning method for sound event classification, which significantly reduces the required amount of labeled samples in the process of classifier training. Active learning is a process of selecting samples to be labeled. In our method, the active learning is based on clustering. We use dictionary-based clustering as the dictionary learning is more suitable to sound event classification. Our classifier will be trained using both unlabelled sound segments (that have predicted labels), and a small number of labeled samples. The proposed method and other reference methods are implemented on a public urban sound dataset with 8732 sound segments, the classification accuracy is used to measure the performance of these classifiers. Experimental results show that the proposed method has higher classification accuracy but requires much less labeled samples than other methods.