VeriSpoof: Beyond Real or Fake with One Audio LLM

Authors: Yang Xiao, Ting Dang

Published: 2026-11-03 10:25:06+00:00

Comment: Tech Report

AI Summary

VeriSpoof is an audio language model that unifies five forensic tasks related to speech deepfakes: detection, generator attribution, recording-condition analysis, synthesized-span localization, and source verification. It provides grounded, verifiable rationales for each decision. Achieving state-of-the-art results among open-weight systems, VeriSpoof offers enhanced transparency and multi-faceted analysis for deepfake forensics.

Abstract

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. We present VeriSpoof, an audio language model that unifies five tasks with shared weights: deepfake detection, generator attribution (generator, acoustic model and vocoder), recording-condition analysis, synthesized-span localisation, and source verification. Each decision is supported by grounded rationales whose statements can be independently verified. Answers are scored by sentence likelihood, allowing detection and source verification to be thresholded without separate heads. Rationale statements can be checked against waveform measurements, span labels, or generator metadata. Counterfactual examples train decisions to follow the stated evidence. Across the 14 Speech-DF-Arena test sets, VeriSpoof-Plus achieves 2.14% pooled EER, ranking third of 28 systems and first among open-weight systems. VeriSpoof-Base, trained only on official training splits of eight corpora, achieves 4.43% EER, less than half that of the next open-weight detector. With the same weights, VeriSpoof-Plus verifies whether recordings from generators unseen in training share a source at 6.2% EER. Controlled studies show that the audio representation governs detection, the LLM governs recording comparison, and source supervision is necessary for neural-codec fakes. Statement-checkable rationales catch partial fakes missed by one-step answers (99.6% vs. 78.4% on PartialEdit) and determine decisions under edits unseen in training (99.5%). Training updates only 29.7M of 2.85B parameters, fitting a single 24 GB GPU with gradient checkpointing. We release both models, inference code, a demo, and VeriSpoof-Bench, the evaluation suite covering these questions.


Key findings
VeriSpoof-Plus achieves 2.14% pooled EER on Speech-DF-Arena, ranking third overall and first among open-weight systems. The same model can verify shared sources from unseen generators with 6.2% EER. Rationales significantly improve accuracy on partial deepfakes (99.6% vs. 78.4% on PartialEdit) and ensure decisions follow stated evidence, even under unseen edits.
Approach
The model casts forensic questions as answer scoring, where the likelihood of a sentence-based answer serves as a score. It employs a Qwen3-ASR encoder and a MiniCPM5-2B LLM, with specific adaptations for audio processing and reasoning. Decisions are supported by rationales with statements checkable against waveform measurements, span labels, or generator metadata, trained with counterfactual examples to ensure faithfulness.
Datasets
ASVspoof 2019 LA, ASVspoof 5, SpoofCeleb, MLAAD, STOPA, DFADD, LibriSeVoc, CodecFake+, PartialSpoof, HAD, PartialEdit, ASVspoof 2015, ADD 2022, Fake-or-Real, ReplayDF, SpeechFake, WaveFake, Codecfake, Common Voice 15.0, LibriSpeech, AISHELL-3, GigaSpeech, CN-Celeb2, Speech-DF-Arena, SpoofRadar v1.0, XMAD-Bench, LlamaPartialSpoof.
Model(s)
Qwen3-ASR encoder with DoRA adapters, MiniCPM5-2B LLM with LoRA adapters.
Author countries
UNKNOWN