About TwineTwine NetworkContact
VoxCeleb
VoxCeleb is a large-scale audio-visual speech dataset built from YouTube interview clips, widely used to train and benchmark deep speaker recognition models for speaker verification, speaker identification, and robust “in-the-wild” voice AI.
Files
1000000
Size
133 MB
Format
MP4
Duration
2,000 + hours. Each file ≥ 3 seconds
Country
Participants
7000
Languages
Updated
December 11, 2025
Description
VoxCeleb is a multimodal dataset of short talking-face speech clips (each ≥ 3 seconds) curated from interview videos, designed for text-independent speaker recognition under realistic conditions like background noise, overlapping speech, laughter, pose variation, and changing lighting. It contains 7,000+ speakers, 1M+ utterances, and 2,000+ hours of audio-video data, making it a go-to benchmark for robustness testing, representation learning, and production-grade speaker embedding evaluation.
Dataset Technical Specification
Number of files:
1000000
Total dataset size:
133 MB
Duration:
2,000 + hours. Each file ≥ 3 seconds
Format:
MP4
Sample rate:
Resolution:
Dataset Demographics
📍 Country:
🧍 Gender:
Male: 61%; Female: 39%
📅 Age:
👥 Number of participants:
7000
