Methods and Datasets for DJ-Mix Reverse Engineering

DJ techniques are an important part of popular music culture. However, they are also not sufficiently investigated by researchers due to the lack of annotated datasets of DJ mixes. Thus, this paper aims at filling this gap by introducing novel methods to automatically deconstruct and annotate recorded mixes for which the constituent tracks are known. A rough alignment first estimates where in the mix each track starts, and which time-stretching factor was applied. Second, a sample-precise alignment is applied to determine the exact offset of each track in the mix. Third, we propose a new method to estimate the cue points and the fade curves which operates in the time-frequency domain to increase its robustness to interference with other tracks. The proposed methods are finally evaluated on our new publicly available DJ-mix dataset UnmixDB. This dataset contains automatically generated beat-synchronous mixes based on freely available music tracks, and the ground truth about the placement, transformations and effects of tracks in a mix.