CITATION — REFERENCE ENTRY
Feature Learning for Chord Recognition: The Deep Chroma Extractor — Proceedings of the 17th International Society for Music Information Retrieval Conference (ISMIR 2016)
- Key
- korzeniowski2016-deep-chroma
- Authors
- Korzeniowski, Filip; Widmer, Gerhard
- Issued
- 2016
- Type
- paper-conference
- Container
- Proceedings of the 17th International Society for Music Information Retrieval Conference (ISMIR 2016)
- Pages
- 37-43
Raw CSL JSON
{
"DOI": "10.5281/zenodo.1416314",
"URL": "https://archives.ismir.net/ismir2016/paper/000178.pdf",
"page": "37-43",
"type": "paper-conference",
"title": "Feature Learning for Chord Recognition: The Deep Chroma Extractor",
"author": [
{
"given": "Filip",
"family": "Korzeniowski"
},
{
"given": "Gerhard",
"family": "Widmer"
}
],
"issued": {
"date-parts": [
[
2016
]
]
},
"event-place": "New York, NY, USA",
"container-title": "Proceedings of the 17th International Society for Music Information Retrieval Conference (ISMIR 2016)"
}
Claims
-
Using chromagrams derived from ground-truth annotations, logistic regression recognized 97% of chords (reduced to major/minor) in the Beatles dataset.
"given chromagrams derived from ground truth annotations, using logistic regression we can recognise 97% of chords (reduced to major/minor) in the Beatles dataset."
-
Korzeniowski and Widmer argued that chroma vectors can in principle carry enough information for chord recognition but that standard extractors produce noisy features, and proposed learning chroma with neural networks.
"We present the argument that chroma vectors potentially hold enough information to model harmonic content of audio for chord recognition, but that standard chroma extractors compute too noisy features. This leads us to propose a learned chroma feature extractor based on artificial neural networks."
-
On 383 songs, the learned chroma reached 78.8% weighted chord symbol recall for major/minor chords, compared with 69.2% for standard constant-Q chroma and 76.1% for the raw quarter-tone spectrogram.
"Our proposed feature extractor CD clearly performs best"
Available in