VeriSpoof: Beyond Real or Fake with One Audio LLM
VeriSpoof-Plus ranked first among open-weight detectors across 14 Speech-DF-Arena test sets, with a pooled equal error rate of 2.14%.
The model shares weights across five tasks: deepfake detection, generator attribution, recording-condition analysis, synthetic-span localisation and source verification. It supplies rationales that can be checked against waveform measurements, span labels or generator metadata. The authors report that these rationales caught partial fakes at 99.6% on PartialEdit, compared with 78.4% for one-step answers. They release both model variants, inference code, a demo and an evaluation suite.
HF Daily Papers' note
The model shares weights across five tasks: deepfake detection, generator attribution, recording-condition analysis, synthetic-span localisation and source verification. It supplies rationales that can be checked against waveform measurements, span labels or generator metadata. The authors report that these rationales caught partial fakes at 99.6% on PartialEdit, compared with 78.4% for one-step answers. They release both model variants, inference code, a demo and an evaluation suite.
HF Daily Papers' note
score 5