VoxCeleb2: 1 Million Utterances for 6,112 Speaker Identities

VoxCeleb2 contains over 1 million audio-visual utterances from 6,112 celebrities, extracted from YouTube videos. This large-scale speaker identification dataset includes MP4 video files and associated metadata for training and development. It was updated in early 2026 by user Oldi451.

Use Cases

• Speaker identification using the identity labels in vox2_meta.csv
• Automatic speech recognition (ASR) using audio extracted from MP4 files
• Audio-visual biometric verification using the video frames and corresponding audio tracks

Strengths

• Over 1 million utterances
• 6,112 unique celebrity identities
• Multimodal data including both video and audio

Limitations

• Potential label noise due to automated extraction from YouTube
• Significant storage requirements for multipart MP4 archives

Provenance

Source

YouTube

Collection Method

Automated extraction from video uploads

Freshness

Last updated January 2026.

Geography

United States