CITATION — REFERENCE ENTRY
20 Years of Automatic Chord Recognition from Audio — Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR 2019)
- Key
- pauwels2019-20-years
- Authors
- Pauwels, Johan; O'Hanlon, Ken; Gómez, Emilia; Sandler, Mark B.
- Issued
- 2019
- Type
- paper-conference
- Container
- Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR 2019)
- Pages
- 54-63
Raw CSL JSON
{
"DOI": "10.5281/zenodo.3527739",
"URL": "https://archives.ismir.net/ismir2019/paper/000004.pdf",
"page": "54-63",
"type": "paper-conference",
"title": "20 Years of Automatic Chord Recognition from Audio",
"author": [
{
"given": "Johan",
"family": "Pauwels"
},
{
"given": "Ken",
"family": "O'Hanlon"
},
{
"given": "Emilia",
"family": "Gómez"
},
{
"given": "Mark B.",
"family": "Sandler"
}
],
"issued": {
"date-parts": [
[
2019
]
]
},
"event-place": "Delft, Netherlands",
"container-title": "Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR 2019)"
}
Claims
-
Most available annotated chord data is English-language pop music, so its representativeness is questionable.
"Since the majority of available annotated data consists of anglophonic pop music, its representativeness is questionable."
-
The five most common chord types account for over 80% of popular-music datasets, an imbalance that affects both training and evaluation.
"the five most common chord types account for over 80% of popular music datasets [6, 64]. This imbalance in chord distribution affects both training data-driven ACR systems and evaluation."
-
Complex chords are still recognized less well than major and minor triads, and results have improved enough that ambiguity in chord labels has become a research topic.
"complex chords are still recognised less well than major and minor triads. Nonetheless, ACR results have improved to the point where the ambiguity in chord labels is now becoming a topic of research."
-
Recent automatic chord recognition research has been dominated by deep learning.
"More recently ACR research has been dominated by Deep Learning (DL)."
-
Fujishima's system distinguished 27 chord types; subsequent systems quickly reduced this to two or four types of triad.
"Fujishima's system was able to distinguish no less than 27 different chord types, but subsequent systems quickly reduced this number to two [2] or four [21] types of triad."
-
Pauwels and colleagues describe Fujishima's 1999 paper as having started an active research topic in and around the ISMIR community.
"This paper kickstarted an active research topic that has been popular in and around the ISMIR community."
-
Pauwels and colleagues identify the move from knowledge-driven to data-driven systems as the most striking change in automatic chord recognition.
"Perhaps the most striking evolution in ACR has been the move from knowledge-driven to data-driven systems."
-
Reported average agreement between annotators on chord roots ranges from 76% (four annotators) to 94% (two annotators), with large outliers in individual files.
"The reported agreement on the root of a chord lies between 76% [29] (4 annotators) and 94% [12] (2 annotators) on average, but large outliers towards the bottom can appear in individual files"
Available in