A brief list of some previous work on chord labelling from MIDI. I may apply my work to this task one day. One of these models has featured as a component in my parser.
Ni et al. is missing from this list
Authors |
Publication |
Models |
Code |
Data |
Notes |
Rhodes, Lewis, Müllensiefen |
Bayesian model selection, Dirichlet distributions |
No |
Not public, but mentioned here |
Makes segmentation decisions by marginalising over labeling distributions and then selects most likely chord label. Most models don't incorporate segmentation. Claim results are comparable to other methods, but easier to incorporate other features into this. Excellent review of other work. |
|
Temperley |
Music and Probability |
Simple key-finding models based on Krumhansl |
? |
Public collections (Essen, Kostka-Payne, etc) |
Could use a Krumhansl-inspired approach like this as an emission model in an HMM or similar to do chord labelling |
HMM generating pitch classes of MIDI data |
No |
No |
Widely cited. Possibly the best candidate for the basis for a category decoder, especially as it predicts functions as well as roots. Unfortunately, no objective results reported. No dataset available (or even described). Train on unlabelled data by initializing with naive chord-type emission assumptions - attractive idea. |
||
Pardo, Birmingham |
Algorithms for Chordal Analysis, 2002 |
Non-probabilistic |
|
Kostka-Payne |
Widely-cited chord labelling technique. Does segmentation. Deliberately simple method to provide a baseline. Evaluates against a gold-standard. |
RaphaelAndStoddard: more detailed description of R&S's model