Submissions[edit]
| Sub Code
|
Extended Abstract
|
Contributors
|
Methods
|
| FZZ1
|
PDF
|
Wanpeng Fan, Jiaye Zhu, Peng Zhong
|
WavLM + Conformer
|
| NUS (baseline)
|
Link
|
Xiaoxue Gao,Chitralekha Gupta, Haizhou Li
|
Genre-informed Silence + Phone Model
|
Results[edit]
Jamendo V1[edit]
Jamendo V1 refers to the 20 English songs in the Jamendo dataset (https://github.com/f90/jamendolyrics) with old annotations. This is the dataset used in previous MIREXes to make a fair comparison with the previous submissions.
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
0.547
|
0.047
|
0.686
|
0.912
|
| NUS
|
0.217
|
0.046
|
0.751
|
0.945
|
Jamendo V2 MultiLang[edit]
Jamendo V2 contains all 79 songs in the Jamendo dataset (https://github.com/f90/jamendolyrics) with new annotations.
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
0.584
|
0.252
|
0.683
|
0.887
|
| NUS
|
0.651
|
0.136
|
0.502
|
0.729
|
Language-Specific Results[edit]
Jamendo V2 En[edit]
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
0.619
|
0.143
|
0.698
|
0.896
|
| NUS
|
0.216
|
0.046
|
0.784
|
0.947
|
Jamendo V2 Fr[edit]
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
0.371
|
0.045
|
0.661
|
0.897
|
| NUS
|
0.809
|
0.157
|
0.400
|
0.665
|
Jamendo V2 Gr[edit]
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
0.371
|
0.045
|
0.661
|
0.897
|
| NUS
|
0.809
|
0.157
|
0.400
|
0.665
|
Jamendo V2 Es[edit]
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
0.452
|
0.039
|
0.703
|
0.905
|
| NUS
|
0.969
|
0.230
|
0.392
|
0.613
|
Hansen's dataset a cappella[edit]
Notice that Hansen's dataset and Mauch's dataset overlap with commonly used training sets (e.g., DALI). The results are shown for reference only.
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
0.101
|
0.044
|
0.783
|
0.971
|
| NUS
|
0.132
|
0.031
|
0.791
|
0.965
|
Hansen's dataset[edit]
Notice that Hansen's dataset and Mauch's dataset overlap with commonly used training sets (e.g., DALI). The results are shown for reference only.
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
3.264
|
3.604
|
0.648
|
0.870
|
| NUS
|
0.107
|
0.052
|
0.764
|
0.972
|
Mauch's dataset[edit]
Notice that Hansen's dataset and Mauch's dataset overlap with commonly used training sets (e.g., DALI). The results are shown for reference only.
| Group
|
Average absolute error
|
Median absolute error
|
Percentage of correct segments
|
Percentage of correct onsets with tolerance
|
| FZZ1
|
0.900
|
0.122
|
0.489
|
0.844
|
| NUS
|
0.192
|
0.098
|
0.478
|
0.910
|
Per-song result[edit]
Please see the CSV file.
We want to thank to previous task captain Georgi Dzhambazov for the support on evaluation scripts and datasets, and Emir Demirel for his suggestions to the task.