VoxCeleb2

2,442-hour large-scale speaker recognition corpus with 6,112 celebrities collected from YouTube across many languages.

• Task: speaker
• Languages: multi
• Hours: 2442
• Domain: celebrity speech (in-the-wild)
• License: see source terms
• Homepage: https://www.robots.ox.ac.uk/~vgg/data/voxceleb/vox2.html
• Paper: https://arxiv.org/abs/1806.05622

Recommendation

The standard pre-training corpus for speaker verification models. Large scale and in-the-wild diversity make it near-mandatory for production speaker embeddings. Academic use only — commercial use requires a separate agreement. Download requires requesting access via the VoxCeleb website.

Getting the data

Obtain from the dataset homepage.

Download requires registration and acceptance of terms at the VoxCeleb website. Audio must be downloaded from YouTube using the provided scripts.

Suggested processing

A recommended VoxKitchen pipeline ships in the repository at voxkitchen/templates/pipelines/speaker-analysis.yaml — run it with vkit docker run.