The MTG-Jamendo Dataset

DOI

We present the MTG-Jamendo Dataset, a new open dataset for music auto-tagging. It is built using music available at Jamendo under Creative Commons licenses and tags provided by content uploaders. The dataset contains over 55,000 full audio tracks with 195 tags from genre, instrument, and mood/theme categories. We provide elaborated data splits for researchers and report the performance of a simple baseline approach on five different sets of tags: genre, instrument, mood/theme, top-50, and overall.

This repository contains metadata, scripts, instructions on how to download and use the dataset and reproduce baseline results.

A subset of the dataset was used in the Emotion and Theme Recognition in Music Task within MediaEval 2019-2021.

Table of contents:

Structure

Metadata files in data

Pre-processing

Subsets

Splits

Note: A few tags are discarded in the splits to guarantee the same list of tags across all splits. For autotagging.tsv, this results in 55,525 tracks annotated by 87 genre tags, 40 instrument tags, and 56 mood/theme tags available in the splits.

Splits are generated from autotagging.tsv, containing all tags. For each split, the related subsets (top50, genre, instrument, mood/theme) are built filtering out unrelated tags and tracks without any tags.

Some additional metadata from Jamendo (artist, album name, track title, release date, track URL) is available in raw.meta.tsv (56,693).

Statistics in stats

Top 20 tags per category

Statistics of number of tracks, albums and artists per tag sorted by number of artists. Each directory has statistics for metadata file with the same name. Here is the statistics for the autotagging set. Statistics for subsets based on categories are not kept seperated due to it already included in autotagging.

Using the dataset

Requirements

The original requirements are kept in reguirements-orig.txt

Downloading the data

All audio is distributed in 320kbps MP3 format. We recommend using this version of audio by default. For smaller download sizes, we also provide a lower-bitrate mono version of the same audio (converted from the full quality version to mono LAME VBR 2 lame -V 2). In addition we provide precomputed mel-spectrograms which are distributed as NumPy Arrays in NPY format (see computation parameters in the code). We also provide precomputed statistical features from Essentia (used in the AcousticBrainz music database) in JSON format. The audio files and the NPY/JSON files are split into folders packed into TAR archives.

We provide the following data subsets:

We provide a script to download and validate all files in the dataset. See its help message for more information:

python3 scripts/download/download.py -h

usage: download.py [-h] [--dataset {raw_30s,autotagging_moodtheme}]
                   [--type {audio,audio-low,melspecs,acousticbrainz}]
                   [--from {mtg,mtg-fast}] [--unpack] [--remove]
                   outputdir

Download the MTG-Jamendo dataset

positional arguments:
  outputdir             directory to store the dataset

options:
  -h, --help            show this help message and exit
  --dataset {raw_30s,autotagging_moodtheme}
                        dataset to download (default: raw_30s)
  --type {audio,audio-low,melspecs,acousticbrainz}
                        type of data to download (audio, audio in low quality,
                        mel-spectrograms, AcousticBrainz features) (default: audio)
  --from {mtg,mtg-fast}
                        download from MTG (server in Spain, slow),
                        or fast MTG mirror (Finland) (default: mtg-fast)
  --unpack              unpack tar archives (default: False)
  --remove              remove tar archives while unpacking one by one (use to
                        save disk space) (default: False)

For example, to download audio for the autotagging_moodtheme.tsv subset, unpack and validate all tar archives:

mkdir /path/to/download
python3 scripts/download/download.py --dataset autotagging_moodtheme --type audio /path/to/download --unpack --remove

Unpacking process is run after tar archive downloads are complete and validated. In the case of download errors, re-run the script to download missing files.

Due to the large size of the dataset, it can be useful to include the --remove flag to save disk space: in this case, tar archive are unpacked and immediately removed one by one.

Loading data in python

Assuming you are working in scripts folder

import commons

input_file = '../data/autotagging.tsv'
tracks, tags, extra = commons.read_file(input_file)

tracks is a dictionary with track_id as key and track data as value:

{
    1376256: {
    'artist_id': 490499,
    'album_id': 161779,
    'path': '56/1376256.mp3',
    'duration': 166.0,
    'tags': [
        'genre---easylistening',
        'genre---downtempo',
        'genre---chillout',
        'mood/theme---commercial',
        'mood/theme---corporate',
        'instrument---piano'
        ],
    'genre': {'chillout', 'downtempo', 'easylistening'},
    'mood/theme': {'commercial', 'corporate'},
    'instrument': {'piano'}
    }
    ...
}

tags contains mapping of tags to track_id:

{
    'genre': {
        'easylistening': {1376256, 1376257, ...},
        'downtempo': {1376256, 1376257, ...},
        ...
    },
    'mood/theme': {...},
    'instrument': {...}
}

extra has information that is useful to format output file, so pass it to write_file if you are using it, otherwise you can just ignore it

Reproduce postprocessing & statistics

Recreate subsets

Reproduce experiments

Results

The MTG-Jamendo Dataset can be linked to related datasets tailored to specific applications.

Music Classification Annotations

The Music Classification Annotations contains annotations for the split-0 test set according to the taxonomies of 15 existing music classification datasets including genres, moods, danceability, voice/instrumental, gender, and tonal/atonal. These labels are suitable for training individual classifiers or learning everything in a multi-label setup (auto-tagging). Most of the taxonomies were annotated by three different annotators. We provide the subset of annotations with perfect inter-annotator agreement ranging from 411 to 8756 tracks depending on the taxonomy.

Song Describer

Song Describer Dataset contains ~1.1k captions for 706 permissively licensed music recordings. It is designed for use in evaluation of models that address music-and-language tasks such as music captioning, text-to-music generation and music-language retrieval. The dataset was built using the Song Describer platform for crowdsourcing music captions (audio-text pairs) for audio tracks in MTG-Jamendo.

Research challenges using the dataset

Citing the dataset

Please consider citing the following publication when using the dataset:

Bogdanov, D., Won M., Tovstogan P., Porter A., & Serra X. (2019). The MTG-Jamendo Dataset for Automatic Music Tagging. Machine Learning for Music Discovery Workshop, International Conference on Machine Learning (ICML 2019).

@conference {bogdanov2019mtg,
    author = "Bogdanov, Dmitry and Won, Minz and Tovstogan, Philip and Porter, Alastair and Serra, Xavier",
    title = "The MTG-Jamendo Dataset for Automatic Music Tagging",
    booktitle = "Machine Learning for Music Discovery Workshop, International Conference on Machine Learning (ICML 2019)",
    year = "2019",
    address = "Long Beach, CA, United States",
    url = "http://hdl.handle.net/10230/42015"
}

License

Copyright 2019-2023 Music Technology Group

Acknowledgments

This work was funded by the predoctoral grant MDM-2015-0502-17-2 from the Spanish Ministry of Economy and Competitiveness linked to the Maria de Maeztu Units of Excellence Programme (MDM-2015-0502).

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 765068.

This work has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 688382 “AudioCommons”.