MelodyShape at MIREX 2014 Symbolic Melodic Similarity

J. Urbano
Music Information Retrieval Evaluation eXchange, 2014.

Best results in the MIREX 2014 Symbolic Melodic Similarity task.


This short paper describes our three submissions to the 2014 edition of the MIREX Symbolic Melodic Similarity task. All three submissions rely on a geometric model that represents melodies as spline curves in the pitch-time plane. The similarity between two melodies is then computed with a sequence alignment algorithm between sequences of spline spans: the more similar the shape of the curves, the more similar the melodies they represent. As in the previous MIREX 2010, 2011, 2012 and 2013 editions, our systems ranked first for all effectiveness measures. The main difference with last year is that we submitted a re-implementation of all algorithms, contained in the new open source library MelodyShape.