CITATION — REFERENCE ENTRY
The Harmonix set: Beats, downbeats, and functional segment annotations of Western popular music — Proceedings of the 20th International Society for Music Information Retrieval Conference
- Key
- nieto2019-harmonix-set
- Authors
- Nieto, Oriol; McCallum, Matthew; Davies, Matthew E. P.; Robertson, Andrew; Stark, Adam; Egozy, Eran
- Issued
- 2019
- Type
- paper-conference
- Container
- Proceedings of the 20th International Society for Music Information Retrieval Conference
- Pages
- 565-572
Raw CSL JSON
{
"URL": "https://archives.ismir.net/ismir2019/paper/000068.pdf",
"page": "565-572",
"type": "paper-conference",
"title": "The Harmonix set: Beats, downbeats, and functional segment annotations of Western popular music",
"author": [
{
"given": "Oriol",
"family": "Nieto"
},
{
"given": "Matthew",
"family": "McCallum"
},
{
"given": "Matthew E. P.",
"family": "Davies"
},
{
"given": "Andrew",
"family": "Robertson"
},
{
"given": "Adam",
"family": "Stark"
},
{
"given": "Eran",
"family": "Egozy"
}
],
"issued": {
"date-parts": [
[
2019
]
]
},
"event-place": "Delft, Netherlands",
"container-title": "Proceedings of the 20th International Society for Music Information Retrieval Conference"
}
Claims
-
The Harmonix Set contains manual beat, downbeat, and functional segment annotations for 912 Western popular music tracks, mostly with stable tempo and 4/4 time; it distributes annotations, metadata, identifiers, and onset times for aligning audio rather than the audio itself.
"The Harmonix set contains manual annotations for 912 western popular music tracks, thus being the largest published dataset to date containing beats, downbeats, and function structural segmentation information. [...] These onsets are intended to help aligning the audio in case researchers obtain audio data with different compression formats [...] As such, most tend to have a very stable tempo and a 4/4 time signature."
-
Difficulties of sharing audio have led to ad hoc use of different beat-tracking datasets in the literature. Datasets differ in whether they contain excerpts or full tracks, number of annotators, annotation post-processing, style specificity, Western or non-Western content, and perceived difficulty. The MIREX downbeat task (from 2014) used six datasets: The Beatles, HJDB, Turkish, Ballroom, Carnatic, and Cretan.
"However, the well-known difficulties of sharing the audio component of large annotated datasets has led to a rather ad-hoc usage of different datasets within the literature [...]. In 2014, the Audio Downbeat Estimation (ADE) task was launched which comprised six different datasets from diverse geographic and stylistic sources: The Beatles [24]; Hardcore, Jungle, Drum and Bass (HJDB) [15]; Turkish [41]; Ballroom [21]; Carnatic [42]; and Cretan [17] [...]. Duration: [...] beat and downbeat tracking datasets can be comprised of musical excerpts [...] rather than full tracks [...]. Number of annotators [...] Annotation post-processing [...] Style-specificity [...] Western / Non-Western [...] Perceived difficulty: Finally, the choice of musical material can be based upon the perceived difficulty of the musical excerpts"
Available in