This is the repository for the paper:

@inproceedings{Ward_2018,
	year = 2018,
	month = {April},
	publisher = {{IEEE}},
	author = {Dominic Ward and Hagen Wierstorf and Russell D. Mason and Emad M. Grais and Mark D. Plumbley},
	title = {{{BSS EVAL or PEASS?} Predicting the Perception of Singing-Voice Separation}},
	booktitle = {2018 {IEEE} International Conference on Acoustics, Speech and Signal Processing ({ICASSP})}
}

What's what?

• The website, hosted on GitHub pages is in the site folder
• The submitted (raw) subjective ratings are in site/_data/results
• Audio are in site/sounds/
• The source code and images for the paper are in the paper folder

The experiment lives in the site folder, which is deployed to the gh-pages branch via

git subtree push --prefix site origin gh-pages

Go to https://cvssp.github.io/perceptual-study-source-separation/ to see the experiment.

Main data files you should care about

There are quite a few intermediate files and things left for future work.

In short, the main data are:

• data/ratings.csv is the compiled subjective dataset
• data/experiment_stimuli.csv describes the main audio files used in the experiment
• data/bss_eval_and_peass_clean.csv the predictions of BSS Eval and PEASS

The stimuli used for the experiment can be found in the ./site/sounds folder. The flac files are provided, but we used original wavs for the lab experiment and all objective measures. If you would like to generate the stimuli yourself, read on.

Python

Create and source the Python 3 virtual environment:

cd ./venvs
make
source py3/bin/active

In order to run the PEASS model, you will need to install the Matlab engine inside the Python virtual environment. To do this, go to $MATLABROOT/extern/engines/python and run:

python setup.py build --build-base=$HOME/tmp/build install

You should now be good to go.

All python scripts should be run from the root folder of the repository.

Generating the stimuli

Datasets

DSD100: The demixing Secret Dataset dataset can be downloaded here. This was used for the ground truth data, i.e. the reference vocals and accompaniments.

MUS2017: For our analysis, we used the SiSEC submission data (~400 GB), which was kindly provided by Fabian-Robert Stöter. You can contact me for the complete submission data.

Scripts

python/generation/generate_familarisation_stimuli.py

Generates the configuration files and audio files for the quality familiarisation page and the interference familiarisation page. You will need to set the path, in the main function, to the demixing Secret Dataset (DSD100) dataset, which can be downloaded here.

python/generation/generate_stimuli.py

This script generates all wav files for the training stages, e.g. here and the main experiment, e.g. here. This includes the reference vocals and original accompaniment (required for objective evaluation), the estimated vocals (from the algorithms) and the anchors. Audio files belonging to a single song are placed in their own folder, e.g. ./site/sounds/vocals-10-SIR holds the audio associated with the vocals of song 10.

The main function requires the path to DSD100 and MUS2017.

Naming convention:

• The estimated files have been named as done here.
• The reference vocal is named ref.flac.
• The accompaniment (sum of other sources) associated with each reference vocal has the name ref_accompaniment.flac.
• Artefacts.wav is the sound-quality anchor.
• Interferer.wav is the original mixture.

Note:

Files with the suffix non_norm were not loudness normalised, nor were they used for the listening test, so we haven't included these in the repository. The purpose of these files was to investigate the sensitivity of PEASS to loudness normalisation, which is not discussed in the paper. Furthermore, files with drums, other or bass in their name were not used (and thus not included) but may be useful for future work.

The reference vocals, estimated vocals and anchors (without non_norm in the filename) were loudness normalised according to ITU-R BS.1770-4. The accompaniment signals, e.g. ref_accompaniment.flac were scaled by the same gain factor used to loudness normalise the reference vocal, i.e.

mixture = gain_used_to_normalise_vocal * ref + gain_used_to_normalise_vocal * ref_accompaniment

In other words, the resulting mixture is just a scaled version of the original mixture, where the loudness of the vocal matches that of the extracted vocals.

Finally, the above script also generates the csv files data/experiment_stimuli.csv and data/training_stimuli.csv.

python/generation/generate_interface_config_file.py

Generates the configuration files required for each instance of the MUSHRA test (training and main experiment).

Subjective Data

python/subjective/compile_dataset.py

Compiles the ratings and stores them as a csv file named ./data/ratings.csv. The python module from rename_subjects is not provided, and was used to preserve the anonymity of a few listeners who entered their name when submitting. You will still be able to run this script, but shouldn't need to.

python/subjective/concordance.py

Plots the concordance coefficients and prints out other measures of agreement (sort of) for each task, based on the replicated trials. This is the script used for the first paragraph of section 3.1 of the paper.

python/subjective/hidden_sounds_swarmplot.py

Generates the Bee Swarm plot shown in Figure 1 of the paper.

python/subjective/kripperndorffs_alpha.py.py

Measures of inter-rater reliability as reported in paragraph 3 of the paper (thanks to kripperndorff-alpha for the python implementation).

Objective Data

python/objective/compute_objective_measures.py

Computes the objective measures according to BSS Eval and PEASS. We are essentially using mir_eval.separation.bss_eval_images for BSS Eval. You will need Matlab to run PEASS; see this repo.

In the main function, you will need to:

• Set the path to the compiled PEASS toolbox
• Specify whether you will be loading wav or flac files. We used wav.

This script generates 3 files:

1. ./data/bss_eval_and_peass.csv

   This is the main data we used for the paper. The predictions are made using the same stimuli as used in the experiment, with the original (reference) vocal and accompaniment (sum of all other instruments) as the ground truth sources. The flac files are included for the accompaniments so you can run this script.

2. ./data/bss_eval_and_peass_all_stems.csv

   We later discovered that you get different results depending on whether you input the vocals, bass, drums and other as separate ground truth sources (rather than the vocal and accompaniment). This is possibly the subject of future work.

3. ./data/bss_eval_and_peass_nonorm_all_stems.csv

   Same as #2 but with no loudness normalisation applied, i.e. it uses the original audio files. This is possibly the subject of future work.

python/objective/clean_measures.py

Transforms the above csv files to long format, applying the task label quality of interferer to the appropriate metric. New files are created with the suffix _clean appended to the filename; ./data/bss_eval_and_peass_clean.csv is the one used in the paper.

python/objective/compare_methods.py

Not used for the paper, but shows differences due to how the ground truth sources are specified (compares ./data/bss_eval_and_peass.csv with ./data/bss_eval_and_peass_all_stems.csv).

python/objective/bee_swarm_plot.py

Computes the within-song Spearman correlations reported in Section 3.2 of the paper, and generates Figure 2.

python/objective/regression_plot.py

Prints the Pearson correlation coefficients and RMSEs as reported in the final paragraph of Section 3 and generates the Figure 3 in the paper.