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.
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.
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-datasetThree 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.
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
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 16Then 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/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.
Below you can find some information about the contents of the dataset and how to load them using Python.
Discogs-VI-20240701.jsonlcorresponds to the Discogs-VI dataset which contains all identified cliques and their metadata. The versions are not matched to Youtube IDs.Discogs-VI-YT-20240701.jsonlcorresponds to Discogs-VI-YT subset, with versions matched to YouTube IDs and with post-processing applied to ensure that each clique has at least two downloaded versions.- However, we could match more videos than we could download in Barcelona between 2023-2024. Depending on your location, maybe you can download more than us.
Discogs-VI-20240701.jsonl.youtube_query_matchedcontains all these YouTube IDs.- Some versions are matched to more than one alternative YouTube ID (1.4 videos per version on average) and the matches are sorted from the highest quality match to the lowest, although all YouTube IDs are official uploads.
Discogs-VI-20240701.jsonlandDiscogs-VI-YT-20240701.jsonlcontain rich metadata and they are large in size (around 7 GB and 4 GB). Therefore we provide a file where only clique, version, and Youtube IDs are provided:Discogs-VI-YT-light-20240701.json. This file is the basis for training neural networks.- We then create train, validation, and test partitions from
Discogs-VI-YT-light-20240701.jsonafter dealing with the test sets of the Da-TACOS and SHS100K datasets (see the paper for more information).Discogs-VI-YT-20240701-light.json.train,Discogs-VI-YT-20240701-light.json.val,Discogs-VI-YT-20240701-light.json.test
discogs_20240701_artists.xml.jsonl.cleancontains detailed artist metadata that may be useful.Discogs-VI-YT-20240701.jsonl.demois to be used with the Streamlit demo for visualization purposes.
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).
discogs_20240701_artists.xml.jsonlis the Discogs artist data dump xml file parsed to a jsonl file with some processing. It contains artist information such as aliases, group memberships, or name variations.discogs_20240701_releases.xml.jsonlis the Discogs release data dump xml file parsed releases to a jsonl file with some processing.discogs_20240701_releases.xml.jsonl.cleanis the cleaned version.discogs_20240701_releases.xml.jsonl.clean.trackscontains the tracks from the clean releases. It is used for identifying the cliques.Discogs-VI-20240701-DaTACOS-SHS100K2_TEST-lost_cliques.txtcontains the clique ids in Discogs-VI that intersect with Da-TACOS and SHS100K test sets.Discogs-VI-20240701.jsonl.queriescontains the query strings that was created to search the versions on YouTube.
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"]:
passDiscogs-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 datawith open("discogs_20240701_artists.xml.jsonl.clean", encoding="utf-8") as infile:
for jsonline in infile:
artist = json.loads(jsonline)discogs_20240701_artists.xml.jsonl,discogs_20240701_artists.xml.jsonl.clean,discogs_20240701_releases.xml.jsonl,discogs_20240701_releases.xml.jsonl.clean, discogs_20240701_releases.xml.jsonl.clean.tracks are JSONL files with utf-8 encoding.Discogs-VI-20240701-DaTACOS-SHS100K2_TEST-lost_cliques.txtandDiscogs-VI-20240701.jsonl.queriesare line-delimited text files.
Please refer to our GitHub Repository for more examples.
Run the demo with Streamlit using:
streamlit run demo.py --server.fileWatcherType -- Discogs-VI-YT-20240701.jsonl.demo
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.
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},
}- The code in this repository is licensed under the Affero GPLv3 license.
- The metadata is licensed under a CC BY-NC-SA 4.0.
- Audio representations are available under request for non-commercial scientific research purposes only.
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.