Ground Truth for Segmentation of Symbolic Music

Music segmentation and structural analysis are crucial issues in Music Information Retrieval. It is, however, difficult to find ground truth, especially for MIDI data. This page contains ground truth for 13 music pieces belonging to differenc genres. Manual segmentation has been performed for all tracks of each music piece seperately, still considering the global structure of each song. The following sections list the songs that are available and explains the structure of the ground truth data.

Songs

The following table gives an overview about the songs that are available:

Songtitle	Filename	Artist	Genre	Number of Tracks
Tango A	ab-astor_tango.mid	Anonymous	Tango	8
Act Naturally	Act_Naturally.mid	Buck Owens	Traditional Country	10
Come Back to Me	comeback.mid	Janet Jackson	Adult Contemporary	12
Dance	DANCE.mid	Anonymous	Techno	4
Down	Down.mid	311	Metal	10
We Have all the Time in the World	louis_armstrong-all_the_time_in_the_world.mid	Louis Armstrong	Smooth Jazz	11
Moving On Up	m_people-moving_on_up.mid	M People	Dance Pop	11
My Girl	my_girl.mid	Otis Redding	Soul	9
Red Wine	redwine.mid	UB40	Reggae	9
Slippin' and Slidin	Slippin_and_Slidin.mid	Little Richard	Rock and Roll	5
Stop in the Name of Love	supremes_stop_in_the_name_of_love.mid	Diana Ross and the Supremes	Soul	13
Tennesse Waltz	Tennessee_Waltz.mid	Roy Acuf	Traditional Country	8
YMCA	YMCA.mid	Village People	Dance Pop	16

All songs were taken from the ​Bodhidharma MIDI data set.

XML Format

The ground truth for each song is given in an XML file named like the corresponding '.mid' file but with extension '.xml'. The following elements are contained in each XML file:

Element	Description	Example
filename	name of the corresponding MIDI file	Act_Naturally.mid
title	title of the composition	Act Naturally
artist	artist of the composition	Buck Owens
nr_tracks	number of tracks within the composition	10
duration_beats	duration of the composition given in beats	460.0
genre	genre the composition belongs to	Traditional Country
tracks	collection of track elements	-
track	single track of the composition	-
tracktitle	title of the track (= title of the composition + ascending number)	Act_Naturally_1
instrument	instrument the track is assigend to (sometimes unknown)	Vocal Layer
segmentation	collection of hierarchical segmentation levels	-
level	collection of segment groups containing non-overlapping segments on one hierarchical level	-
segment_group	collection of similar segments	-
segment	occurence of a repeating pattern within a segment group	-
start_note	first note within the segment (simultaneous notes count as one)	134
start_beat	beat where the segment starts	247.5
end_note	last note within the segment (simultaneous notes count as one)	265
end_beat	beat where the segment ends	435.0

All 13 XML files can be downloaded here​.