Reducing manual labeling in singing voice detection: An active learning approach
Wei Li, Xiangyi Feng, Min Xue · 2016
In order to solve the bottleneck of tedious and time-consuming manual labeling in singing voice detection, in this paper we integrate the active learning mechanism into the conventional SVM-based supervised learning algorithm. By selecting most informative unlabeled samples and asking for human annotation, active learning substantially reduces the number of training samples to be labeled and meanwhile obtains almost the same frame-level vocal/non-vocal classification performance. Experiments on the public MIR-1K database demonstrate the effectiveness of active learning in the task of singing voice detection.