Compressing Music Recordings into Audio Summaries

We present a criterion to generate audible summaries of music recordings that optimally explain a given track with mutually disjoint segments of itself. We represent audio as sequences of beat-synchronous harmonic features and use an exhaustive search to identify the best summary. To demonstrate the merit of this approach, we evaluate the criterion and show consistency across a collection of multiple recordings of different works. Finally, we present a fast algorithm that approximates the exhaustive search and allows us to automatically learn the hyperparameters of the algorithm for a given track.