2010:Harmonic Analysis

From MIREX Wiki
Revision as of 01:06, 14 May 2010 by Sutashu (talk | contribs)

Description

This task is suggested for MIREX 2010. Generally speaking, the goal is to estimate the latent harmonic progression from score-like music data, a process typically called harmonic analysis or functional analysis in musicology. This typically involves estimating the musical key of a given piece of music and movement of the tonal center (modulations and tonicizations), as well as the progression of chords.

Data format

Not much data exists that can be used as ground truth for harmonic analysis. Several researchers used Bach chorales ([6] and [8]) and some reported using Kostka-Payne corpus. Unfortunately, the latter is very small, too small in fact for proper model training and evaluation: it consists of 46 very short musical excerpts.

Input data

We propose to use a dataset developed in cooperation of the members of Sagayama/Ono laboratory at the University of Tokyo and prof. Hitomi Kaneko from Toho Gakuen School of Music and her students. This dataset contains very detailed harmony labels for all of RWC database's classical pieces (almost 6 hours of music). This data will be coupled with quantized (score-like) set of MIDI files, and the latter will be converted to CSV files for easy input. The first line of such CSV file will contain information about the number of data lines per downbeat and the number of lines per upbeat (Auftakt). The rest of the lines will hold lists of notes active in the consecutive divisions of downbeat expressed by MIDI note numbers. The first two bars of Chopin's Nocturne Op. 9 No. 2 (see picture above) will therefore be encoded as:

4 1
62
39 55 58 63 67 79
51 56 59 62 68 77 79
39 55 58 63 67 77
38 55 58 63 67 70 75
...

In other words, there would be 4 lines per measure, each spanning 3 eighth notes and the first line will be a 3 eighth notes of upbeat.

Ground truth

The ground truth for the same piece will be encoded as:

z
E- E- G B-
B- E- A- D F
E- E- G B-
E- D E- G B-
...

Each line will consist of harmony data corresponding to one line of the input data. Harmony is given by pitch class of the root tone followed by a list of all pitch classes belonging to the chord sorted by the frequencies of notes found in the data. Pitch classes are labeled with capital Latin letters (A through G) with flats represented by minus signs and sharps by plus signs. No harmony (e.g. a rest or upbeat) is marked with 'z'. So, in the example above, “E- E- G B-” means a chord with root in E♭ (which is the tonic) that consists of three notes: E♭, G and B♭, and so it is a major triad in its root position.

The tonal center (and therefore the structural meaning of a given chord) is not explicitly given in order to simplify the task and allow more participants to compete. Nevertheless, the algorithms will benefit much from considering e.g. tonal center movement and degrees of scale.

Labels for the entire RWC's classical portion will be made available for the MIREX participants (with some copying and usage restrictions). In order to allow cross-validation of the submitted algorithms, an independent set of labels will be created (to which the participants will have no access), and the algorithms will be executed on both sets.

Output data

The output from the algorithms is expected to comply with the above description of the ground truth data format. For example, a common 24-chord dictionary algorithm would output a series of lines from the following set (12 major and 12 minor chords):

C C E G
C+ C+ E+ G+
D- D- F A-
D D F+ A
D+ D+ F A+
E- E- G B-
E E G+ B
F F A C
F+ F+ A+ C+
G- G- B- D-
G G B D
G+ G+ B+ D+
A- A- C E-
A A C+ E
A+ A+ C E+
B- B- D F
B B D+ F+
C C E- G
C+ C+ E G+
D- D- F- A-
D D F A
D+ D+ F+ A+
E- E- G- B-
E E G B
F F A- C
F+ F+ A C+
G- G- B D-
G G B- D
G+ G+ B D+
A- A- C- E-
A A C E
A+ A+ C+ E+
B- B- D- F
B B D F+

Evaluation Measures

References

  1. F. Lerdahl, “Tonal Pitch Space,” Oxford University Press, 2001
  2. W. B. de Haas, R. C. Veltkamp, F. Wiering, “Tonal pitch step distance: a similarity measure for chord progressions,” Proc. of 9th ISMIR, 2008
  3. J.-F. Paiement, D. Eck, S. Bengio, “Chord representations for probabilistic models,” IDIAP Research report, 2005
  4. J.-F. Paiement, D. Eck, S. Bengio, ”A probabilistic model for chord progressions,” Proc. of 6th ISMIR, 2005
  5. C. Raphael, J. Stoddard, “Harmonic analysis with probabilistic graphical models,” Proc. of 4th ISMIR, 2003
  6. P. Kröger, A. Passos, M. Sampaio, G. de Cidra, “Rameau: a system for automatic harmonic analysis,” Proc. of ICMC, 2008
  7. C. S. Sapp, “Computational chord-root identification in symbolic musical data: rationale, methods and applications,” Computing in Musicology 15, 2007
  8. H. Taube, “Automatic tonal analysis: toward the implementation of a music theory workbench,” Computer Music Journal vol. 23 nr. 4, 1999


Potential Participants

Discussion for 2010