RESULTS

In this page, you can find all the plots used in the paper plus others.

Raw results

First of all, here are the raw results for each exeriment. Each file contains several raw in this format:

Piece number: [x]
Precision: [x]
Recall: [x]
F1-measure: [x]

The piece number is the index [starting from 0] in the alphabetically sorted list of files of the directory relative to the experiment.

Crossvalidation on Pop dataset

• Chew voice-separation (not cross)
• Skyline (not cross)
• CNN non mono
• CNN mono

Cross-validation on Mozart dataset

• Chew voice-separation (not cross)
• Skyline (not cross)
• CNN non mono
• CNN mono

Validation on cultivated music

• Chew voice-separation
• Skyline
• CNN non mono
• CNN mono
• CNN non mono
• CNN mono

Plots

N.B. You can recreate these plots by using the Jupyter Notebook provided in this repo.

You can download the high quality SVG plots by clicking on the images. The p-value refers to the Friedman test (for more details about the significance see the full paper).

Results of the cross-validation on the Pop song dataset. The plot shows precision, recall and F-measure of all the predictions from all the folds.

Results of the cross-validation on the Mozart song dataset. The plot shows precision, recall and F-measure of all the predictions from all the folds.

Results of the Validation on the cultivated dataset. The plot shows precision, recall and F-measure of all music pieces contained in the dataset.

Results of the Validation on the cultivated dataset. The plot shows precision, recall and F-measure averaged on each author. The numbers between parenthesis near the authors is the number of songs per each author.