Four Timely Insights on Automatic Chord Estimation — Proceedings of the 16th International Society for Music Information Retrieval Conference (ISMIR 2015)

Claims

1. billboard-size
   The McGill Billboard dataset consists of over 1,000 annotations, of which more than 700 have been made public.

   "The McGill Billboard dataset consists of over 1000 annotations, of which more than 700 have been made public."

   Locator: page: 674 · Quote language: en
2. class-imbalance
   Likelihood scaling after training was used to control class bias caused by the severe imbalance in the distribution of chords.

   "likelihood scaling is performed after training to control class bias resulting from the severe imbalance in the distribution of chords."

   Locator: page: 674 · Quote language: en
3. cnn-constant-q
   Humphrey and Bello's convolutional network takes roughly one-second patches of constant-Q spectra as input.

   "Time-frequency patches of local contrast normalized constant-Q spectra, on the order of one second, are transformed by a four-layer convolutional network."

   Locator: page: 674 · Quote language: en
4. four-insights
   Humphrey and Bello conclude that not all music content is valid for chord estimation; that conventional comparison methods struggle with the relationships between chords; that methodology mixes the goals of transcription and recognition; and that the subjectivity of chord perception may make objective ground truth untenable.

   "one, not all music content is valid in the context of chord estimation; two, conventional comparison methods struggle to accurately characterize the complex relationships between chords; three, conventional methodology has mixed the somewhat conflicting goals of chord transcription and recognition to an undefined degree; and four, the subjective nature of chord perception may render objective ground truth and evaluation untenable."

   Locator: page: 678 · Quote language: en
5. i-saw-her
   For the Beatles song 'I Saw Her Standing There', which has several independent reference annotations, no two references agreed on more than 65% at the tetrad level.

   "no two reference annotations correspond to greater than a 65% agreement"

   Locator: page: 677 · Quote language: en
6. isophonics
   The first major reference chord annotation effort, now part of the Isophonics dataset, covers the 180-song Beatles discography plus 20 songs by Queen, 14 by Carole King and 18 by Zweieck.

   "covers the entire 180-song discography of The Beatles, as well as 20 songs from Queen, 14 from Carole King, and 18 from Zweieck"

   Locator: section: 2.3.1 · Quote language: en
7. pipeline
   Most chord estimation research converged on a common architecture: chroma features extracted from short frames, optional pre-filtering, pattern matching against chord classes, and post-filtering into a sequence of chord labels.

   "the bulk of chord estimation research has concentrated on building better systems, mostly converging to a common architecture [4]: first, harmonic features, referred to as pitch class profiles (PCP) or chroma, are extracted from short-time observations of the audio signal [7]; these features may then be processed by any number of means, referred to in the literature as pre-filtering;"

   Locator: page: 673 · Quote language: en
8. vocabulary-157
   Humphrey and Bello's vocabulary comprised the 13 most frequent chord qualities on all 12 roots plus a no-chord class, 157 classes in all.

   "comprised of the thirteen most frequent chord qualities in all twelve pitch classes, as well as a no-chord class, for a total of 157 chord classes"

   Locator: page: 675 · Quote language: en