It might come as a surprise that behind an artistic task such as djing lie many computational challenges. In practical terms, the input to the problem is given by a set of "songs" (tracks), and the output consists in an uninterrupted stream of music obtained by suitably adjusting and overlapping a subset of the input tracks. In order to solve this problem both automatically and with results that are qualitatively comparable those of a professional dj, one has to solve tasks such as beat tracking, tempo estimation, music structure analysis, and mix quality evaluation, just to name a few. The solutions for all of these tasks build upon techniques from digital signal processing (in particular FFTs) and machine learning (neural networks). In this talk, we first establish a set of rules that a satisfactory solution --a Dj mix-- has to satisfy, and then give an overview of the algorithmic techniques used to solve the aforementioned problems. Finally, we present our results after one year of work.
