VeriSpoof audio LLM unifies deepfake detection and source verification
Original title: VeriSpoof: Beyond Real or Fake with One Audio LLM
VeriSpoof unifies five tasks with shared weights and releases open models achieving 2.14% EER on Speech-DF-Arena.
VeriSpoof is an audio language model that performs deepfake detection, generator attribution, recording-condition analysis, synthesized-span localisation and source verification using shared weights and grounded rationales. It achieves 2.14% pooled EER on 14 test sets and 4.43% EER for the Base variant. Models are released along with inference code and a demo for single-GPU training.
- VeriSpoof-Plus ranks third of 28 systems
- Base model less than half next open-weight EER
- Verifies unseen generators at 6.2% EER
- Statement-checkable rationales catch partial fakes
- Trains by updating 29.7M of 2.85B parameters
- AI models
- VeriSpoof
From the original article
1 min read at huggingface.co
A 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, Ting Dang · huggingface.co
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.