VeriSpoof audio LLM unifies deepfake detection and source verification
Original title: VeriSpoof: Beyond Real or Fake with One Audio LLM
VeriSpoof achieves 2.14% EER for deepfake detection and generator verification across 14 test sets
VeriSpoof is an audio language model that performs five unified tasks including deepfake detection, generator attribution, recording-condition analysis, synthesized-span localisation, and source verification. It achieves 2.14% pooled EER ranking third among 28 systems and first among open-weight systems. The model supports counterfactual examples and is trained with shared weights for grounded rationales.
- Verifies source for unseen generators at 6.2% EER
- Training updates only 29.7M of 2.85B parameters
- Statement-checkable rationales catch partial fakes missed by one-step answers
- Released models, code, and VeriSpoof-Bench
From the original article
1 min read at arxiv.orgA speech deepfake detector estimates whether a recording is synthetic, but a forensic analyst also asks which system produced it, whether recordings share a source, and what evidence supports each conclusion.
By Yang Xiao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.