Landmark-based Audio Fingerprinting for DJ Mix Monitoring

Recently, the media monitoring industry shows increased interest in applying automated audio identification systems for revenue distribution of DJ performances played in dis- cotheques. DJ mixes incorporate a wide variety of signal modifications, e.g. pitch shifting, tempo modifications, cross-fading and beat-matching. These signal modifica- tions are expected to be more severe than what is usually encountered in the monitoring of radio and TV broadcasts. The monitoring of DJ mixes presents a hard challenge for automated music identification systems, which need to be robust to various signal modifications while maintaining a high level of specificity to avoid false revenue assignment. In this work we assess the fitness of three landmark-based audio fingerprinting systems with different properties on real-world data – DJ mixes that were performed in dis- cotheques. To enable the research community to evaluate systems on DJ mixes, we also create and publish a freely available, creative-commons licensed dataset of DJ mixes along with their reference tracks and song-border annota- tions. Experiments on these datasets reveal that a recent quad-based method achieves considerably higher perfor- mance on this task than the other methods.