CN-Celeb

1,200-hour multi-genre Mandarin speaker recognition corpus spanning 11 real-world scenarios collected from Chinese celebrities.

• Task: speaker
• Languages: zh
• Hours: 1200
• Domain: celebrity speech (in-the-wild)
• License: see source terms
• Homepage: https://www.openslr.org/82
• Paper: https://arxiv.org/abs/1911.01799

Recommendation

The primary benchmark for Mandarin speaker verification and identification in realistic, in-the-wild conditions. Wide acoustic diversity across genres (interview, singing, entertainment) is valuable but also makes it challenging. Check the source terms carefully — the data is free for research but redistribution is restricted.

Getting the data

Downloadable via VoxKitchen (cnceleb, source: openslr, size: 20.7 GB):

vkit docker download --tag slim cnceleb --root ./data/cnceleb

Subsets: cn-celeb_v2.

Suggested processing

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