2008:Audio Chord Detection Results

From MIREX Wiki

Introduction

These are the results for the 2008 running of the Audio Chord Detection task set. For background information about this task set please refer to the 2008:Audio Chord Detection page.

Task Descriptions

Task 1 (Pretrained Systems) Go to Task 1 Results: Systems were pretrained and they were tested against 176 Beatles songs.

Task 2 (Train-Test Systems) Go to Task 2 Results: System trained on ~2/3 of the beatles dataset and tested on ~1/3. Album filtering was applied on each train-test fold such that the songs from the same album can not appear in both train and test sets simultaneously.

Overlap score was calculated as the ratio between the overlap of the ground truth and detected chords and ground truth duration. Also a secondary overlap score was calculated by ignoring the major-minor variations of the detected chord (e.g., C major == C minor, etc.).

Note that 4 songs were excluded from the original Beatles dataset because of alignment of ground truth to the audio problems. The ground truth to audio alignment was done automatically. The script to perform the alignment is going to be released soon by Chris Harte.

General Legend

Team ID for ChordPreTrained (Task 1)

BP = J. P. Bello, J. Pickens
KO = M. Khadkevich, M. Omologo
KL1 = K. Lee 1
KL2 = K. Lee 2
MM = M. Mehnert
PP = H.Papadopoulos, G. Peeters
PVM = J. Pauwels, M. Varewyck, J-P. Martens
RK = M. Ryynänen, A. Klapuri

Team ID for ChordTrainTest (Task 2)

DE = D. Ellis
ZL = X. Jhang, C. Lash
KO = M. Khadkevich, M. Omologo
KL = K. Lee (withtrain)
UMS = Y. Uchiyama, K. Miyamoto, S. Sagayama
WD1 = J. Weil
WD2 = J. Weil, J-L. Durrieu

Overall Summary Results

Task 1 Results

Task 1 Overall Results

<csv>2008/chord/task1_results/pretrained_summary.csv</csv>

<csv>2008/chord/task1_results/pretrained_runtimes.csv</csv>

Task 1 Summary Data for Download

File Name Set (Pretrained runs)
Summary Overlap Data (Pretrained runs)
Summary Overlap Data (Pretrained runs (Merged maj/min))

Task 1 Friedman's Test for Significant Differences

The Friedman test was run in MATLAB against the Task 1 Overlap Score data over the 176 ground truth songs.


<csv>2008/chord/task1_results/task1_friedman.csv</csv>

The Tukey-Kramer HSD multi-comparison data below was generate using the following MATLAB instruction.

Command: [c,m,h,gnames] = multcompare(stats, 'ctype', 'tukey-kramer', 'estimate', 'friedman', 'alpha', 0.05); <csv>2008/chord/task1_results/friedman_detailed.csv</csv>

2008 task1.friedman.png

Task 2 Results

Task 2 Overall Results

<csv>2008/chord/task2_results/summary.csv</csv>

<csv>2008/chord/task2_results/task2_runtimes.csv</csv>

Task 2 Summary Data for Download

File Name Set (Train-test runs)
Summary Overlap Data (Train-test runs)
Summary Overlap Data (Train-Test runs (Merged maj/min))
Per Fold Summary Data (Train-Test runs (Zip archive))

Task Friedman's Test for Significant Differences

The Friedman test was run in MATLAB against the Task 2 Overlap Score data over the 176 ground truth songs.


<csv>2008/chord/task2_results/all3folds_friedman.txt</csv>

The Tukey-Kramer HSD multi-comparison data below was generate using the following MATLAB instruction.

Command: [c,m,h,gnames] = multcompare(stats, 'ctype', 'tukey-kramer', 'estimate', 'friedman', 'alpha', 0.05); <csv>2008/chord/task2_results/task2_allFolds_friedman_detailed.csv</csv>

2008 task2.allfolds friedman.png