VeriSpoof: Beyond Real or Fake with One Audio LLM

VeriSpoof. Beyond Real or Fake with One Audio LLM. By Yang Xiao and one other, from the University of Melbourne. Published on the third of November.

A speech deepfake detector estimates one thing: whether a recording is synthetic. A forensic analyst needs more than that single answer. Which system produced the recording. Whether two recordings share a source. And what evidence supports each of those conclusions, because a conclusion nobody can check is hard to rely on. A detector that only says real or fake leaves every one of those questions open.

VeriSpoof is an audio language model that handles five tasks with one set of shared weights. Deepfake detection. Generator attribution, which goes down to the generator, the acoustic model and the vocoder. Recording-condition analysis. Localisation of synthesized spans, meaning which parts of a recording were generated. And source verification. Rather than five separate tools, it is one model asked five kinds of question. Each decision is supported by grounded rationales, made of statements that can be independently verified.

Answers are scored by sentence likelihood. That lets detection and source verification be thresholded like any detector, without separate heads bolted on to produce a score. The rationale statements are built to be checkable: against waveform measurements, against span labels, or against generator metadata. Counterfactual examples are used in training so that decisions follow the evidence the model has stated, rather than sitting beside it. The training is light. It updates only twenty-nine point seven million of two point eight five billion parameters, which fits on a single twenty-four gigabyte GPU with gradient checkpointing.

On the fourteen Speech-DF-Arena test sets, VeriSpoof-Plus reaches a pooled equal error rate of two point one four percent. That ranks third of twenty-eight systems, and first among open-weight systems. VeriSpoof-Base, trained only on the official training splits of eight corpora, reaches four point four three percent, less than half the error of the next open-weight detector. With the same weights, VeriSpoof-Plus verifies whether recordings from generators unseen in training share a source, at six point two percent error.

Controlled studies take the system apart. The audio representation governs detection, the LLM governs comparison between recordings, and source supervision is necessary for neural-codec fakes. The checkable rationales matter on their own too. On PartialEdit, they catch partial fakes that one-step answers miss: ninety-nine point six percent against seventy-eight point four percent. Under edits unseen in training, they determine the decisions, at ninety-nine point five percent.

The shift is from a verdict to a case. Instead of a single real-or-fake score, the model takes on the analyst's further questions, and backs its answers with statements that can each be checked against the audio. VeriSpoof-Bench, the evaluation suite covering these questions, is released alongside both models.