Collecting ground truth annotations for drum detection in polyphonic music.pdf

COLLECTING GROUND TRUTH ANNOTATIONS FOR DRUM DETECTION IN POLYPHONIC MUSIC Koen Tanghe1, Micheline Lesaffre1, Sven Degroeve2, Marc Leman1, Bernard De Baets2 and Jean-Pierre Martens4 1 IPEM, Department of Musicology, Ghent University, Blandijnberg 2, 9000 Ghent, Belgium Koen.Tanghe@UGent.be 2 Department of Applied Mathematics, Biometrics and Process Control, Ghent University 4 Department of Electronics and Information Systems, Ghent University ABSTRACT In order to train and test algorithms that can automati- cally detect drum events in polyphonic music, ground truth data is needed. This paper describes a setup used for gathering manual annotations for 49 real-world mu- sic fragments containing different drum event types. Apart from the drum events, the beat was also anno- tated. The annotators were experienced drummers or percussionists. This paper is primarily aimed towards other drum detection researchers, but might also be of interest to others dealing with automatic music analysis, manual annotation and data gathering. Its purpose is threefold: providing annotation data for algorithm train- ing and evaluation, describing a practical way of setting up a drum annotation task, and reporting issues that came up during the annotation sessions while at the same time providing some thoughts on important points that could be taken into account when setting up similar tasks in the future. Keywords: drum detection, annotation, data gathering 1 INTRODUCTION Drum events provide important clues about the rhythmical organisation of a musical piece. For many music genres nowadays, rhythmic structures have be- come (at least) equally important as melodic or tonal structures. In the same way that melody lines can be seen as a representation of one aspect of the musical content of a piece of music, drum sequences can be seen as another type of musical content representation, but more related to rhythm. This information could be used to allow p