Download VoxCeleb Dataset

Introduction

Most of the download scripts on the website are not able to access the VoxCeleb datasets, so I decide to share how I did it in this article.

Request Access

Fill the form below to request access to the VoxCeleb datasets.

Download Script

VoxCeleb 1

Replace <USER> and PASSWORD with your own.

USER=<USER>
PASSWORD=<PASSWORD>

wget http://cnode01.mm.kaist.ac.kr/voxceleb/vox1a/vox1_dev_wav_partaa --user $USER --password $PASSWORD
wget http://cnode01.mm.kaist.ac.kr/voxceleb/vox1a/vox1_dev_wav_partab --user $USER --password $PASSWORD
wget http://cnode01.mm.kaist.ac.kr/voxceleb/vox1a/vox1_dev_wav_partac --user $USER --password $PASSWORD
wget http://cnode01.mm.kaist.ac.kr/voxceleb/vox1a/vox1_dev_wav_partad --user $USER --password $PASSWORD

wget http://cnode01.mm.kaist.ac.kr/voxceleb/vox1a/vox1_test_wav.zip --user $USER --password $PASSWORD

cat vox1_dev_wav* > vox1_dev_wav.zip
unzip vox1_dev_wav.zip

VoxCeleb 2

Replace <USER> and PASSWORD with your own.

USER=<USER>
PASSWORD=<PASSWORD>

wget http://143.248.230.30/voxceleb/vox1a/vox2_dev_aac_partaa --user $USER --password $PASSWORD
wget http://143.248.230.30/voxceleb/vox1a/vox2_dev_aac_partab --user $USER --password $PASSWORD
wget http://143.248.230.30/voxceleb/vox1a/vox2_dev_aac_partac --user $USER --password $PASSWORD
wget http://143.248.230.30/voxceleb/vox1a/vox2_dev_aac_partad --user $USER --password $PASSWORD
wget http://143.248.230.30/voxceleb/vox1a/vox2_dev_aac_partae --user $USER --password $PASSWORD
wget http://143.248.230.30/voxceleb/vox1a/vox2_dev_aac_partaf --user $USER --password $PASSWORD
wget http://143.248.230.30/voxceleb/vox1a/vox2_dev_aac_partag --user $USER --password $PASSWORD
wget http://143.248.230.30/voxceleb/vox1a/vox2_dev_aac_partah --user $USER --password $PASSWORD

wget http://143.248.230.30/voxceleb/vox1a/vox2_test_aac.zip --user $USER --password $PASSWORD

cat vox2_dev_aac* > vox2_dev_aac.zip
unzip vox2_dev_aac.zip

For VoxCeleb 2 dataset, we need to convert flac to wav format. I dug into this rabbit hole and provided the Python script to do so. Just simply replace <VOX2_DEV_ROOT> and <VOX2_TEST_ROOT> with your own path.

import os
from glob import glob
from tqdm.auto import tqdm


def main():
    dev_root = '<VOX2_DEV_ROOT>'
    for f in tqdm(glob(os.path.join(dev_root, '**/*/*.m4a')), leave=False):
        f_in = f
        f_out = f.replace('.m4a', '.wav')
        os.system(f'ffmpeg -i {f_in} {f_out}')
        os.system(f'rm {f_in}')
    test_root = '<VOX2_TEST_ROOT>'
    for f in tqdm(glob(os.path.join(test_root, '**/*/*.m4a')), leave=False):
        f_in = f
        f_out = f.replace('.m4a', '.wav')
        os.system(f'ffmpeg -i {f_in} {f_out}')
        os.system(f'rm {f_in}')


if __name__ == '__main__':
    main()

It can also be adapted to a faster way using multi-thread with tqdm library.

import os
from glob import glob
from tqdm.auto import tqdm


def convert(filepath):
    f_in = filepath
    f_out = filepath.replace('.m4a', '.wav')
    os.system(f'ffmpeg -i {f_in} {f_out}')
    os.system(f'rm {f_in}')


def main():
    dev_root = '<VOX2_DEV_ROOT>'
    process_map(
        convert, 
        glob(os.path.join(dev_root, '**/*/*.m4a')), 
        max_workers=8, 
        chunksize=1024
    )

    test_root = '<VOX2_TEST_ROOT>'
    process_map(
        convert, 
        glob(os.path.join(dev_root, '**/*/*.m4a')), 
        max_workers=8, 
        chunksize=1024
    )


if __name__ == '__main__':
    main()

Extra Datasets

1. TIMIT
2. MUSAN
3. RIR
4. VCTK