Song2Quartet: A System for Generating String Quartet Cover Songs from Polyphonic Audio of Popular Music

We present Song2Quartet, a system for generating string quartet versions of popular songs by combining probabilistic models estimated from a corpus of symbolic classical music with the target audio file of any song. Song2Quartet allows users to add novelty to listening experience of their favorite songs and gain familiarity with string quartets. Previous work in automatic arrangement of music only used symbolic scores to achieve a particular musical style; our challenge is to also consider audio features of the target popular song. In addition to typical audio music content analysis such as beat and chord estimation, we also use timefrequency spectral analysis in order to better reflect partial phrases of the song in its cover version. Song2Quartet produces a probabilistic network of possible musical notes at every sixteenth note for each accompanying instrument of the quartet by combining beats, chords, and spectrogram from the target song with Markov chains estimated from our corpora of quartet music. As a result, the musical score of the cover version can be generated by finding the optimal paths through these networks. We show that the generated results follow the conventions of classical string quartet music while retaining some partial phrases and chord voicings from the target audio.