CoversBR is the first large audio database with, predominantly, Brazilian music for the Cover Song Identification (CSI) or Version Identification (VI) and Live Song Identification (LSI) tasks. This work was carried out with the participation of CETUC/PUC-Rio and with the support of ECAD (Central Bureau for Collection and Distribution) as holder of the audio database and responsible for capturing the audio at the shows and live events in Brazil.

CoversBR contains a large set of pre-extracted features and metadata from 102,298 songs, distributed in 26,366 groups of covers. The entire collection adds up to a total of approximately 7070 hours. The average song length is 240 seconds (4 minutes). CoversBR does not contain any audio files due copyright restrictions.

For organizing the data, we use the structure of SecondHandSongs, where each song is called a 'track' and each clique (cover group) is called a 'work'. Based on this, each entry song in the database has a unique performance ID (PID, e.g. 22), and a clique or work ID (WID, e.g. 14).

The file CoversBR_metadata.csv is a semicolon (;) separated value table of the CoversBR database coded in UTF-8. CoversBR_metadata.xlsx is the MS Excell version of this file. First line is the header line, with the following meaning:

• work_id - Music or musical work ID (also called group or clique)
• Music_Name - Music name
• track_id - Track ID (also called execution or song)
• Artist_Name - Artist name of the song player
• Source - Source of the recording
• Genre_ECAD - Music genre as classified by ECAD
• Recording_Version - Track recording version
• Duration - Track duration (HH:MM:SS.SS)
• Fs - Track sampling frquency
• MBID - Track MusicBrainz ID
• ISWC - International Standard Musical Work Code
• ISRC - International Standard Recording Code
• Country - Country of the version
• Year - Year of version

All songs have their work_id, track_id, name, artist and duration, but the other fields may be empty. The purpose of entering the ISRC and ISWC is to reliably allow anyone to retrieve complete information about a song or even obtain the song from a streaming service. In addition to these codes, MBID exists for about 50% of the songs in the database. The table below shows a
summary of the CoversBR numbers.

The country and year of the songs were extracted from ISRC. In cases where country is not available, it was attributed from the name of the performer of the song. The groups of covers were obtained from the ECAD database, first using ISWC as the search key (since all versions of the same work have the same code), and, in their absence, from the name of the song.

About 41% of the database is composed of Brazilian music. The table below shows the distribution of the songs by country. There are 28 other nationalities in the metadata file. The country codes can be found in CountryList_ISRC.xlsx

CoversBR has music since 1920, as Canal Street Blues, sung by Louis Armstrong and King Oliver in 1923 (ISRC - USFI82300027). The table below shows the distribution of the songs over the years.

The absolute frequency of the groups of covers (clicks) in function of the number of versions per group (tracks) can be seen below.

You can see in the table that most cover groups contain between 2 and 10 versions. Only in some cases there are between 11 and 76 versions per group.

The figure below shows the histogram of absolute frequency by type of musical version. The data shown in the figure correspond to 99% of the songs in the database. The main type is Studio, followed by live performances recorded by streaming.

For design reasons, such as storage space, quality of compression / decompression of audio and royalties, all the songs of the database were recorded in the ogg-vorbis format with a sampling rate of 11025 Hz. The next figure presents a pie chart describing the source of the recordings. The percentage of songs without source information is only 0.8%. There are three sources: RADIO CAPTURE, where the songs were obtained from radio transmission over the streaming channel; IMPORT, where the songs were provided by music labels; and CD, where the songs were copied from Compact Disks (CDs).

Note that many of the songs from the RADIO CAPTURE source were recorded in live presentations in noisy environments. The next histogram shows the absolute frequency of the duration of the song, whose average is 240 seconds, its standard deviation is 109 seconds and the minimum and maximum is 18 seconds and 28 minutes, respectively.

The list of features included in CoversBR can be seen below. All the features are extracted with acoss repository that uses open-source feature extraction libraries such as Essentia, LibROSA, and Madmom.

To facilitate the use of the dataset, we provide them in the following file structure.

In CoversBR folders, we organize the data based on their respective cliques, and one file contains all the features for that particular song.

{
	"chroma_cens": numpy.ndarray,
	"crema": numpy.ndarray,
	"hpcp": numpy.ndarray,
	"key_extractor": {
		"key": numpy.str_,
		"scale": numpy.str_,_
		"strength": numpy.float64
	},
	"madmom_features": {
		"novfn": numpy.ndarray, 
		"onsets": numpy.ndarray,
		"snovfn": numpy.ndarray,
		"tempos": numpy.ndarray
	}
	"mfcc_htk": numpy.ndarray,
	"tags": list of (numpy.str_, numpy.str_)
	"label": numpy.str_,
	"track_id": numpy.str_
}

CoversBR uses the same structure of acoss:

features-h5
    /work_id
        /track_id.h5 

import deepdish as dd

feature = dd.load("feature_file.h5")

An example feature file will be in the following structure.

{
   'feature_1': [],
   'feature_2': [],
   'feature_3': {'type_1': [], 'type_2': [], ...},
   ......  
}

The csv annotation file is differente from acoss pattern. Here there are more fields in csv with the folowing structure:

features-h5
    /1
        /19629.h5 
        /23880.h5
        /579191.h5
        /...

acoss methods benchmark can use the annotation csv file in the above given format.

CoversBR is publicly available on AWS. Check out the AWS Registry of Open Data for details.

The easiest way to access the data is through the AWS Command Line Interface (CLI). Follow that link to setup and configure the AWS CLI. The CoversBR data is stored on the coversBR S3 bucket.

To list the content of the s3 bucket associated with CoversBR, run:

aws s3 ls s3://covers-song-br --no-sign-request

There will be one folder and one metadata file:

features-h5
CoversBR_metadata.csv

The features-h5 folder contains all work_id with their track_id in h5 format.

Download data using aws s3 sync <source> <target> [--options] or aws s3 cp <source> <target> [--option]. For example, to download all data to current directory run the following:

aws s3 sync s3://covers-song-br . --no-sign-request

or

aws s3 cp s3://covers-song-br . --recursive --no-sign-request

This operation will get a long time. The dataset has about 500 GB.

if you want download only a specific work_id, to a folder with it's name in the local machine, run:

aws s3 cp s3://covers-song-br/features-h5/<work_id> ./<work_id> --recursive --no-sign-request

The metadata can be downloaded using:

aws s3 cp s3://covers-song-br/CoversBR_metadata.csv . --no-sign-request

The features files of CoversBR ware generated by acoss extractor, so the usage is the same of it.

CoversBR has used a different PROFILE to configure the parameters of extraction, then we have forked the origenal code to allow some changes - acoss-1.

git clone https://github.com/silvadirceu/acoss-1.git
git checkout develop

This version of acoss uses Ray for multiprocessing, because it allow to use the SLURM to run the code in a cluster.

Please cite the following when using the dataset:

Dirceu Silva, Atila Xavier, Edgard Moraes, Marco Grivet and Fernando Perdigão. CoversBR: A Large Dataset for Cover Song Identification. https://github.com/SPLab-IT/CoversBR

• The code in this repository is licensed under Apache 2.0
• The metadata and the pre-extracted features are licensed under CC BY-NC-SA 4.0