Towards Automatic Content-Based Separation of DJ Mixes into Single Tracks

DJ mixes and radio show recordings constitute an important and underexploited music and data source. In this paper we try to approach the problem of separation of a continuous DJ mix into single tracks or timestamping a mix. Sharing some aspects with the task of structural segmentation, this problem has a number of distinctive features that make difficulties for structural segmentation algorithms designed to work with a single track. We use the information derived from spectrum data to separate tracks from each other. We show that the metadata that usually comes with DJ mixes can be exploited to improve the separation. An iterative algorithm that can consider both content-based data and user provided metadata is proposed and evaluated on a collection of freely available timestamped DJ mix recordings of various styles.