The dataset viewer is not available for this subset.
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Provenance-separated 34data/asvspoof2021-df-eval
Repackages 34data/asvspoof2021-df-eval at revision 1ebfec00f42068f3c032460c9694ff541ad61bd1 without decoding or re-encoding audio. Original filenames and audio bytes are preserved. Use the upstream dataset's applicable license; repackaging grants no additional rights.
real/ contains recorded human speech and classical channel transformations. synthetic/ contains generated speech and full neural codec waveform reconstructions. ASVspoof 2021 LA/DF use official bona-fide/spoof protocol labels. For the ASVspoof 5 bona-fide mirror, C04/C07 neural codec reconstructions are assigned synthetic; other conditions remain real. These are provenance conventions, which differ from anti-spoofing task labels.
Counts: {'synthetic': 146460, 'real': 5500}. provenance.json records pinned sources, protocol references, class counts, and shard hashes. Every distinct source filename matched the reviewed protocol and appears exactly once across the output shards. Repeated filenames are deduplicated only after full SHA-256 byte equality checks; conflicting repetitions abort publication. All output archives passed CRC verification.
- Downloads last month
- 183