CITATION — REFERENCE ENTRY
Large-Vocabulary Chord Transcription via Chord Structure Decomposition — Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR 2019)
- Key
- jiang2019-chord-decomposition
- Authors
- Jiang, Junyan; Chen, Ke; Li, Wei; Xia, Gus
- Issued
- 2019
- Type
- paper-conference
- Container
- Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR 2019)
- Pages
- 644-651
Raw CSL JSON
{
"DOI": "10.5281/zenodo.3527892",
"URL": "https://archives.ismir.net/ismir2019/paper/000078.pdf",
"page": "644-651",
"type": "paper-conference",
"title": "Large-Vocabulary Chord Transcription via Chord Structure Decomposition",
"author": [
{
"given": "Junyan",
"family": "Jiang"
},
{
"given": "Ke",
"family": "Chen"
},
{
"given": "Wei",
"family": "Li"
},
{
"given": "Gus",
"family": "Xia"
}
],
"issued": {
"date-parts": [
[
2019
]
]
},
"event-place": "Delft, Netherlands",
"container-title": "Proceedings of the 20th International Society for Music Information Retrieval Conference (ISMIR 2019)"
}
Claims
-
The Billboard dataset contains 230 different chord qualities, or 2,749 distinct chord classes.
"contains 230 different chord qualities, or equivalently, 2,749 distinct chord classes"
-
Jiang and colleagues decompose each chord label into components such as triad, bass and seventh, each with a much smaller vocabulary, recognize them with a multitask classifier, and reassemble the full label.
"The core concept is to decompose any chord label into a set of musically meaningful components (e.g., triad, bass, seventh), each with a much smaller vocabulary compared to the size of the overall chord vocabulary."
-
Classifying each frame independently tends to produce excessive chord changes, because there is no transition penalty between frames.
"this approach tends to produce excessive chord changes as there is no transition penalty between two frames."
-
In the Billboard dataset, the most frequent 10% of chord qualities cover 93.86% of the data, while the least frequent half cover only 0.35%.
"While the first 10% chord qualities cover 93.86% of the data, the last 50% chord qualities only cover 0.35% of the data altogether"
Available in