Measured openly.
We measure on public test sets used by the entire research field, with the Speech DF Arena evaluation: one model, sets measured in full, and we say what the table does not show.
One model, eight sets.
Equal Error Rate (EER) per set: the point where missed attacks and false alarms are equally frequent. Lower is better. Own measurement of our current detection model on the complete test sets, as of October 2026.
All values come from the same model, measured on the complete test sets with the official Speech DF Arena evaluation. We selected the model on our own validation data. We have not submitted it, so there is no ranking.
Training contained no test or development data from these sets, but it did include the public training portions of ASVspoof 2019 (also the basis of the 2021 sets), ASVspoof 5 and DFADD. For InTheWild, SONAR and Fake or Real there was no training share.
The Arena also scores LibriSeVoc, Codecfake and four ADD sets. We have not yet measured LibriSeVoc on the complete corpus. Codecfake and the ADD sets are licensed for non-commercial use only, so we show no values for them. On several of these sets the error rate is higher than in the table.
Test sets and evaluation: Speech DF Arena, Hugging FaceWhat these values mean, and what they do not.
Lab against lab
Benchmark sets are controlled conditions: known attacks, clean cuts, defined channels. The values say how well a method holds against the state of research.
Your daily reality is harder
In production, voices are compressed, multilingual, with echo and background noise. No benchmark mirrors your telephony exactly. Anyone claiming otherwise measures wrong.
Pilot on your material
That is why every deployment starts with your recordings: we measure on your channels, your languages, your noise, and set the thresholds together.
The only number that counts: yours.
The archive screening shows how VeriGate performs on your recordings: with findings per case and a risk profile of your call traffic.