discogs-vi-dataset

Discogs-VI Dataset

DOI

Discogs-VI is a dataset of musical version metadata and pre-computed audio representations, created for research on version identification (VI), also referred to as cover song identification (CSI). It was created using editorial metadata from the public Discogs music database by identifying version relationships among millions of tracks, utilizing metadata matching based on artist and writer credits as well as track title metadata. The identified versions comprise the Discogs-VI dataset, with a large portion of it mapped to official music uploads on YouTube, resulting in the Discogs-VI-YT subset.

In the VI literature the set of tracks that are versions of each other is defined as a clique. Here’s an example of the metadata for a clique. Discogs-VI contains approximately 1.9 million versions belonging to around 348,000 cliques, while Discogs-VI-YT includes approximately 493,000 versions across about 98,000 cliques.

This website accompanies the dataset and the related publication, providing summary information, instructions on access and usage, as well as the code to re-create the dataset, including audio downloads from the matched YouTube videos. The code for dataset re-creation can be found here.

Table of contents

Discogs

Discogs regularly releases public data dumps containing comprehensive release metadata (such as artists, genres, styles, labels, release year, and country). See an example of a release page. See how the Discogs database is built here. You can see some statistics for all music releases submitted to Discogs on their explore page.

Dependencies

We use Python 3.10.9 on Linux.

git clone https://github.com/MTG/discogs-vi-dataset
cd discogs-vi-dataset
conda env create -f environment.yaml
conda activate discogs-vi-dataset

Download

Three types of data are associated with the dataset: clique metadata (Discogs-VI), clique metadata with YouTube ID-matched versions (Discogs-VI-YT), and audio representations such as CQT (Constant-Q Transform) extracted for the versions of Discogs-VI. This section provides details on how to access each type of data.

Metadata

We provide the dataset including the intermediary files of the creation process. Due to their sizes, they are separated into two directories so that you do not have to download everything. If your goal is to use the dataset and start working, download main.zip (1.4 GB compressed, 21 GB uncompressed). If for some reason you are interested in the intermediary files, download intermediary.zip (8.7 GB compressed, 46 GB uncompressed). Contents of these folders are provided in this section. You can download the data from Zenodo

Audio

You can download the audio files corresponding to the YouTube IDs of the versions. In our experiments, we used exactly these IDs.

We have been able to conduct the downloads from our research institution under Directive (EU) 2019/790 on Copyright in the Digital Single Market, which includes text and data mining exceptions for the purposes of scientific research (Article 3).

python discogs_vi_yt/audio_download_yt/download_missing_version_youtube_urls.py Discogs-VI-YT-20240701.jsonl music_dir/

However, Discogs-VI-20240701.jsonl.youtube_query_matched contains more versions with YouTube IDs (read the paper for understanding why or check this section).

python discogs_vi_yt/audio_download_yt/download_missing_version_youtube_urls.py Discogs-VI-20240701.jsonl.youtube_query_matched music_dir/

NOTE: We recommend parallelizing this operation because there are many audio files using utilities/shuffle_and_split.sh. However, if you use too many parallel processes you may get banned from YouTube. We experimented with 2-20 processes. Using more than 10 processes got us banned a few times. In that case, you should stop downloading and wait a couple of days before trying again.

utilities/shuffle_and_split.sh Discogs-VI-YT-20240701.jsonl 16

Then open up multiple terminal instances and call each split separately.

python discogs_vi_yt/audio_download_yt/download_missing_version_youtube_urls.py Discogs-VI-20240701.jsonl.youtube_query_matched.split.00 music_dir/

Audio representations

This repository does not contain the code for extracting the CQT audio representations used to train the Discogs-VINet described in the paper, nor the features themselves. The model and code to extract the features are available in a separate repository. The extracted features are available upon request for non-commercial scientific research purposes. Please contact Music Technology Group to make a request.

Data Structure

Below you can find some information about the contents of the dataset and how to load them using Python.

Main files

NOTE: Every clique and version has a unique ID associated to them. Currently the clique IDs change between Discogs dumps (will be fixed in the code later).

Intermediary files

Loading with python

The files have different encodings and structure. Here you can find how to load each file.

Discogs-VI-20240701.jsonl, Discogs-VI-20240701.jsonl.youtube_query_matched, and Discogs-VI-YT-20240701.jsonl

# Read the file with utf-8 encoding
with open("Discogs-VI-YT-20240701.jsonl", encoding="utf-8") as in_f:
    # Read the file one line at a time
    for jsonline in in_f:
        # Load the clique
        clique = json.loads(jsonline)
        # Access the versions
        for version in clique["versions"]:
            # Access the urls or other metadata. For Discogs-VI-20240701.jsonl there are no youtube_video field
            for video in version["youtube_video"]:
                pass

Discogs-VI-YT-20240701-light.json, Discogs-VI-YT-20240701-light.json.train, Discogs-VI-YT-20240701-light.json.val, and Discogs-VI-YT-20240701-light.json.test

# Read the file with default encoding
with open("Discogs-VI-YT-light-20240701.json") as in_f:
    # Load the cliques
    cliques = json.load(in_f)
# Access the data

Rest of the files

with open("discogs_20240701_artists.xml.jsonl.clean", encoding="utf-8") as infile:
    for jsonline in infile:
        artist = json.loads(jsonline)

Please refer to our GitHub Repository for more examples.

Discogs-VI-YT Streamlit demo

Run the demo with Streamlit using:

streamlit run demo.py --server.fileWatcherType -- Discogs-VI-YT-20240701.jsonl.demo

Snapshot of the interface

Re-create the dataset

The steps to re-create the dataset is detailed in a separate README file. Since Discogs database is growing one can run the scripts periodically and extend the dataset. We plan to create a new version of the dataset every year or so.

Cite

Please cite the following publication when using the dataset:

R. O. Araz, X. Serra, and D. Bogdanov, “Discogs-VI: A musical version identification dataset based on public editorial metadata,” in Proceedings of the 25th International Society for Music Information Retrieval Conference (ISMIR), 2024.

@inproceedings{araz_discogs-vi_2024,
 title = {Discogs-{VI}: {A} musical version identification dataset based on public editorial metadata},
 booktitle = {Proceedings of the 25th {International} {Society} for {Music} {Information} {Retrieval} {Conference} ({ISMIR})},
 author = {Araz, R. Oguz and Serra, Xavier and Bogdanov, Dmitry},
 year = {2024},
}

License

Acknowledgements

This work is supported by “IA y Música: Cátedra en Inteligencia Artificial y Música” (TSI-100929-2023-1) funded by the Secretaría de Estado de Digitalización e Inteligencia Artificial and the European Union-Next Generation EU, under the program Cátedras ENIA.

This site is open source. Improve this page.