ASVspoof 2021 DF — speaker-disjoint 60/10/30 protocol split

ASVspoof 2021 DF — speaker-disjoint 60/10/30 protocol split Three protocol files that partition the ASVspoof 2021 DF evaluation set into speaker-disjoint train / dev / eval subsets, so that a trainable experiment can be run on a corpus that ships as evaluation-only. No audio is included, altered or redistributed. These are 5-column text files referencing the official ASVspoof 2021 filenames. Obtain the audio from the official release and apply these protocols to it: Audio: https://zenodo.org/records/4835108 — DOI 10.5281/zenodo.4835108, open access, four ASVspoof2021_DF_eval_part0{0..3}.tar.gz parts, about 34.5 GB, released under ODC-ODbL. Keys and meta-labels are not in that record; the official copy is at https://www.asvspoof.org/index2021.html. You do not need them here — the KEY and SYSTEM_ID columns carry that information. Why this exists Canonical ASVspoof 2021 DF is eval-only: one set of 611,829 trials with no train/dev partition. Work that needs a trainable variant has to invent a split, and an invented split that is not published is an unreproducible result. This is ours, published so the numbers computed on it can be checked. Contents file trials speakers ASVspoof2021_DF.train.txt 372,456 55 ASVspoof2021_DF.dev.txt 83,871 10 ASVspoof2021_DF.eval.txt 155,502 28 total 611,829 93 611,829 trials is the canonical DF evaluation set exactly — every trial is assigned to exactly one subset, none is added and none is dropped. The class balance is 22,617 bona fide / 589,212 spoof, also the canonical composition. Properties, verified on regeneration Speaker-disjoint: zero speaker overlap between any pair of subsets. This is the property the split exists for — a random trial-level split would leak speakers across subsets and inflate results. Complete and non-duplicating: 611,829 unique filenames, no repeats. Deterministic: regenerating from source twice, independently, produced byte-identical files (matching sha256 sums). Note the split is speaker-disjoint, not label-stratified, so the bona fide proportion varies by subset (3.6% train, 2.8% dev, 4.5% eval). That follows from partitioning on speakers. Format Five space-separated columns, matching the ASVspoof protocol convention: SPEAKER_ID FILENAME SYSTEM_ID - KEY SPEAKER_ID and SYSTEM_ID are carried over unchanged from the official ASVspoof 2021 DF metadata. SYSTEM_ID is - for bona fide trials; for spoof trials it is the source-system tag: A07–A19 for the ASVspoof 2019 LA attacks, and HUB-*, SPO-*, Task1-team*, Task2-team* for the Voice Conversion Challenge 2018 and 2020 systems. KEY is bonafide or spoof. The fourth column is a literal -, retained for positional compatibility with the official protocol files. LA_0023 DF_E_2000011 A14 - spoof Integrity SHA256SUMS.txt covers every other file in this record. Verify with: shasum -a 256 -c SHA256SUMS.txt Every property claimed above — the counts, the speaker disjointness, the completeness, the class balance — is checkable from these files alone, with no audio download and no network access. README.md gives the exact shell commands. The files are LF-terminated. Do not let a tool rewrite the line endings — it changes every checksum. Licence Open Data Commons Open Database License (ODC-ODbL) v1.0, matching the upstream corpus. These files are a partition of the ASVspoof 2021 DF database, which is itself released under ODC-ODbL. Because they extract and re-arrange that database's contents — filenames, speaker identifiers, attack codes and keys — they are best treated as a Derivative Database under ODbL §4.4, which must be licensed under ODbL in turn. We adopt that reading deliberately rather than claiming the files are a mere Produced Work, so that no one inherits a licence question from us. In practice this means: use them freely, including commercially; if you publish a modified or extended version of this database, publish it under ODbL too, and keep the attribution notice. The underlying audio is not included here and is not ours to relicense. It remains under the terms you accept from the official release. Citation Please cite both this record and the corpus it partitions: Yamagishi, J. et al. ASVspoof 2021: Accelerating Progress in Spoofed and Deepfake Speech Detection. Proc. ASVspoof Challenge Workshop, 2021.