Mining Labeled Data from Web-Scale Collections for Vocal Activity Detection in Music

This work demonstrates an approach to generating strongly labeled data for vocal activity detection by pairing instrumental versions of songs with their original mixes. Though such pairs are rare, we find ample instances in a massive music collection for training deep convolutional networks at this task, achieving state of the art performance with a fraction of the human effort required previously. Our error analysis reveals two notable insights: imperfect systems may exhibit better temporal precision than human annotators, and should be used to accelerate annotation; and, machine learning from mined data can reveal subtle biases in the data source, leading to a better understanding of the problem itself. We also discuss future directions for the design and evolution of benchmarking datasets to rigorously evaluate AI systems.