VeriSpoof finds voice deepfakes at 2.14% error with one model

Researchers released VeriSpoof, an open audio language model that finds voice deepfakes and tells which system made the fake and which parts are fake. VeriSpoof-Plus, one of 2 models that the authors release, gets a pooled equal error rate of 2.14% on 14 test sets. At this rate, incorrect alarms and fakes that it does not find are equal.

How it works

VeriSpoof gives a rationale with each decision. A reader can examine each rationale against sound measurements, labels of spans or data of the generator. The training data has examples in which the decision must follow the facts in the rationale.

The numbers

• VeriSpoof-Base has less than half the error of the next open weight detector.
• For fakes that change only a part of a recording, rationale checks get 99.6%, against 78.4% for one step answers.
• For generators that are new to the model, the source check error is 6.2%.

Training changes a small part of the model and uses one GPU.

What is not known

These are the results of the authors. The authors give no test on live calls or on audio with noise.

Insight: Detector scores depend on the test, so demand rationales

In October 2026, one study showed that the procedure that selects a checkpoint can move the score of an image detector by 5.34 points. VeriSpoof gives rationales that a reader can examine and gives results on new generators. This reads as a move from one yes or no score to proof that a buyer can audit. A detector with no rationale is hard to trust.