Voxceleb1 dataset Download

Preface

VoxCeleb1 Is widely used Speaker recognition 、 verification Data sets . Because it is from YouTube Extract from video , There is rich noise .（ Make up the introduction when you are free ）

If you can use Google forms and translation software, you should be able to download smoothly , Distributing datasets privately risks infringement .

Text

The official website is as follows ：

VoxCelebhttps://www.robots.ox.ac.uk/~vgg/data/voxceleb/

But what's amazing is now （2022-7-12）, All download links to this website have been cancelled .

VoxCelebhttps://www.robots.ox.ac.uk/~vgg/data/voxceleb/vox1.html

You can see , It can only be downloaded to Metadata, Audio files are temporarily unavailable .

After searching for a long time, I found that the link below can be downloaded , At first, I was worried that it was not the official website , Later, it was found that this was a South Korean Laboratory , Undertook the fourth VoxCeleb Speaker Recognition Challenge (VoxSRC) 

VoxCelebhttps://mm.kaist.ac.kr/datasets/voxceleb/ Before downloading, you need to fill in Google Form, Fill in the name of the unit . Because it is an automatic process , So you can check your email inbox soon after filling it out , You will see an email giving User name and password .

https://docs.google.com/forms/d/e/1FAIpQLSdQhpq2Be2CktaPhuadUMU7ZDJoQuRlFlzNO45xO-drWQ0AXA/viewform?fbzx=7440236747203254000

Here are instructions , The given identity can only be used 1 Months .

It's easy to get the user name and password , Use Windows And the browser can directly find the corresponding data set download in the following link , Because it's too big, the official made it into pieces , The specific operation is detailed on the official website , Click the link when downloading , You need to fill in the user name and password , Enter to start downloading .VoxCelebhttps://mm.kaist.ac.kr/datasets/voxceleb/

Add something extra ,Linux Download command of environment .

wget http://cnode01.mm.kaist.ac.kr/voxceleb/vox1a/vox1_test_wav.zip  --http-user=username--http-passwd=password

take link `http://cnode01.mm.kaist.ac.kr/voxceleb/vox1a/vox1_test_wavip` Switch to the file you need to download , then username and password Just replace .

The official website gives md5, You can check it easily .

 md5sum vox1_dev_wav.zip

Then decompression , use unzip command .

unzip -d vox1_dev_wav vox1_dev_wav.zip

Then the big work was done , The use of data sets can refer to GitHub look for voxceleb trainer, In addition to using Pytorch Users of can refer to  torchaudio.datasets.voxceleb1 — Torchaudio nightly documentation. This API Relatively new , The older version may not have .

Add

For the data set to be used Train The model student added ,Identification The training of tasks should also be downloaded Test Data .

Direct use  https://mm.kaist.ac.kr/datasets/voxceleb/meta/iden_split.txt  This file reads the data set , Will report a mistake ,id10270-id10309 The data of is missing , however iden_split This document is marked with some id The data of speakers in this range is Training, I thought it was just Training Data （ Because it's not doing ASV） So I didn't download Test…… It turned out to be a mistake , Audio file not found .