Streaming from MIDI Using Constraint Satisfaction Optimization and Sequence Alignment

We present a new system for separating streams in musical pieces encoded as MIDI files. Our approach is to: (1) divide the music under analysis into short segments, (2) analyze each segment using constraint satisfaction optimization, and (3) connect these analyses using a sequence alignment algorithm. Parameters for the system are learned automatically on a small training corpus and generalize reasonably well across a variety of pieces. We report performance results on 108 pieces of Baroque, Classical, and Romantic music: J.S. Bach’s two-part inventions (0.95 accuracy by the F-measure), three-part sinfonias (0.92), and fugues from the Well-Tempered Clavier, book I (0.93) and book II (0.92); and string quartets by Haydn (0.81) and Brahms (0.76).