Datasets:

Modalities:
Audio
Languages:
English
ArXiv:
License:
moisesdb / README.md
vinismarques's picture
Fix license_name for HuggingFace
3162378
metadata
language:
  - en
pretty_name: MoisesDB
tags:
  - audio
  - music
  - source separation
license: other
license_name: cc-by-nc-sa-4.0
license_link: https://creativecommons.org/licenses/by-nc-sa/4.0/

MoisesDB

Moises Dataset for Source Separation

Dataset Description

Dataset Summary

MoisesDB is a dataset for source separation. It provides a collection of tracks and their separated stems (vocals, bass, drums, etc.). The dataset is used to evaluate the performance of source separation algorithms.

Download the data

Please download the dataset at our research website, extract it and configure the environment variable MOISESDB_PATH accordingly.

export MOISESDB_PATH=./moises-db-data

The directory structure should be

moisesdb:
    moisesdb_v0.1
        track uuid 0
        track uuid 1
        .
        .
        .

Install

You can install this package with

pip install git+https://github.com/moises-ai/moises-db.git

Usage

MoisesDB

After downloading and configuring the path for the dataset, you can create an instance of MoisesDB to access the tracks. You can also provide the dataset path with the data_path argument.

from moisesdb.dataset import MoisesDB

db = MoisesDB(
    data_path='./moisesdb',
    sample_rate=44100
)

The MoisesDB object has iterator properties that you can use to access all files within the dataset.

n_songs = len(db)
track = db[0]  # Returns a MoisesDBTrack object

MoisesDBTrack

The MoisesDBTrack object holds information about a track in the dataset, perform on-the-fly mixing for stems and multiple sources within a stem.

You can access all the stems and mixture from the stem and audio properties. The stem property returns a dictionary whith available stems as keys and nd.array on values. The audio property results in a nd.array with the mixture.

track = db[0]
stems = track.stems  # stems = {'vocals': ..., 'bass': ..., ...}
mixture track.audio # mixture = nd.array

The MoisesDBTrack object also contains other non-audio information from the track such as:

  • track.id
  • track.provider
  • track.artist
  • track.name
  • track.genre
  • track.sources
  • track.bleedings
  • track.activity

The stems and mixture are computed on-the-fly. You can create a stems-only version of the dataset using the save_stems method of the MoisesDBTrack.

track = db[0]
path =  './moises-db-stems/0'
track.save_stems(path)

Performance Evaluation

We run a few source separation algorithms as well as oracle methods to evaluate the performance of each track of the MoisesDB. These results are located in csv files at the benchmark folder.

Citing

If you used the MoisesDB dataset on your research, please cite the following paper.

@misc{pereira2023moisesdb,
      title={Moisesdb: A dataset for source separation beyond 4-stems}, 
      author={Igor Pereira and Felipe Araújo and Filip Korzeniowski and Richard Vogl},
      year={2023},
      eprint={2307.15913},
      archivePrefix={arXiv},
      primaryClass={cs.SD}
}

Licensing

MoisesDB is distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

For the complete license terms, please visit: https://creativecommons.org/licenses/by-nc-sa/4.0/

See LICENSE file for details.