ASVspoof 2021: 54 teams' spoofed and deepfake speech detectors across logical access, physical access and deepfake tasks

The record is the second voice entry in the studies index, two editions before ASVspoof 5. No product sold under a card in this index is a named participant.

Participants, not products, are the detectors here: 54 teams submitted valid scores and 34 the required system description, from 72 registered; teams are identified only by number, and unlabelled rows in the result tables sent no description (arXiv 2210.02437, Section VII, Acknowledgements and Fig. 3 caption, read 10 November 2026). Four organiser baselines were scored in every task.

Ranks from Tables XIV to XVI and descriptions from Table V of arXiv 2210.02437.

None of the voice cards in this index, Pindrop Pulse, Resemble Detect, Reality Defender, Hive, Sensity, ElevenLabs or AI or Not, appears as a participant. Deepfake-O-Meter hosts research implementations of RawNet2 and LFCC-LCNN trained on ASVspoof data, architectures shared with the baselines, not scored challenge entries. The paper cites a Pindrop blog post as background and prints no Pindrop result (arXiv 2210.02437, reference 6, read 10 November 2026).

All data are English read speech from the VCTK corpus. The LA task passes the ASVspoof 2019 LA attacks through real telephony in seven conditions, from uncoded 16 kHz audio through a-law, G.722, mu-law, GSM and OPUS codecs over VoIP and an uncontrolled PSTN leg, with three conditions unseen before evaluation. The PA task replays speech in nine real rooms. The DF task, new in 2021, adds bona fide and spoofed speech from the 2018 and 2020 Voice Conversion Challenges, compressed with mp3, m4a and ogg at two bit rates each plus two double-codec conditions (arXiv 2210.02437, Section II and Tables I to IV, read 10 November 2026). Training used only the 2019 data, so the VCC sources were unseen.

The LA and PA metric is the pooled normalised minimum t-DCF, which scores the detector in tandem with a fixed speaker verification system, with EER also printed; the DF task has no verification system and is scored by EER alone (arXiv 2210.02437, Section II, read 10 November 2026).

Every figure below is printed in Tables XIV to XVI of the supplementary material or in Tables VI, VIII and X of the paper.

Values as printed in arXiv 2210.02437, Tables VI, VIII, X and XIV to XVI, read 10 November 2026. Table VIII prints T23's LA min t-DCF as 0.2176 and Table VI its DF EER as 15.63%, against 0.2177 and 15.64% in the supplementary tables.

Three findings carry beyond the leaderboard. First, telephony did little harm in the LA task: transmission across real telephone systems "causes only modest degradation", though the voice conversion attack A18 rose from a median t-DCF of about 0.4 uncoded to over 0.8 after the unknown PSTN leg (arXiv 2210.02437, Sections III-A and VII, read 10 November 2026). Second, the DF task exposed a source mismatch rather than a codec problem: EERs for the VCC source data were "substantially higher" than for ASVspoof 2019 data, and when VCC bona fide trials were excluded T23's EER fell below 1%, which the authors trace to a shift in bona fide score distributions and "over-fitting" to the 2019 training corpus (arXiv 2210.02437, Section IV-C and Table VI, read 10 November 2026).

Third, silence was a cue. On hidden subsets with non-speech intervals removed, top LA and DF systems that were under 10% EER rose above 20%, and trimming only the leading and trailing silence took B04's LA min t-DCF from 0.426 to 0.908; four of the top ten PA systems scored above 0.99 on a simulated replay subset, which the authors say cautions against simulated evaluation data (arXiv 2210.02437, Sections IV-A and IV-B, Table X and supplementary Section X, read 10 November 2026).

Section VI lists them: the data are "exclusively English-language, read-speech sourced from the VCTK corpus"; the LA and DF attacks were state of the art "prior to 2020"; the strict training policy excluded the external pre-training that post-challenge systems used to reach 0.82% LA EER; non-speech length "should not serve as a cue for detection"; only fully spoofed utterances were scored, not partial spoofs; and adversarial attacks on the detector were out of scope (arXiv 2210.02437, Section VI, read 10 November 2026). The DF task also lacked a human-listener leg, which the authors suggest for future tandem assessment.

Evident from the paper: per-team results sit in the supplementary tables, with the main text showing bar charts; 20 of the 54 teams are unlabelled; and nothing identifies a commercial product.

No card takes a result line from this record, because no indexed product is a scored participant. The record counts toward the voice column of studies by modality as a record with zero indexed detectors, and the content type view lists the cards that state voice. No figure enters the false positive rate ranking, which ranks cards. The home page lists which AI detectors exist and who has tested them.

Related records: ASVspoof 5 is the 2024 edition, where the best open-condition detector reached 2.59% EER and again posted 10% to 27% on other databases; SAFE 2025 is the index's other audio challenge, with named teams; DFDC 2020 is the video counterpart.