RAT: Reference-Augmented Training for ASV Anti-Spoofing

A training strategy that conditions an anti-spoofing model on a speaker reference recording — then learns to ignore it — has delivered state-of-the-art deepfake detection results on the ASVspoof 5 benchmark, according to a preprint posted to arXiv [1]. The approach, called Reference-Augmented Training (RAT), was introduced in a paper submitted to the open-access repository on June 9, 2026 [1]. The authors designed a spoofing countermeasure architecture that receives a reference recording of the claimed speaker alongside the test utterance. During training, the model converges to a solution that effectively disregards the reference channel, yet the process induces an invariance that strengthens detection [1]. “Surprisingly, training with a reference channel induces invariance that improves deepfake detection, even when the reference is absent or mismatched during inference,” the researchers write [1]. On the ASVspoof 5 benchmark, a single RAT detector achieved a 2.57% equal error rate and a minimum detection cost function of 0.074, surpassing even large ensemble systems [1]. The optimization process rapidly diminishes the reference contributions, so at inference the model can operate with a zero vector in place of a real reference recording without performance loss [1]. The paper appears on arXiv, a repository that hosts electronic preprints across physics, computer science, and related fields and has grown to a submission rate of about 24,000 articles per month as of November 2024 [6]. Unlike journal publications, arXiv e-prints are moderated but not peer-reviewed [6]. The RAT manuscript is listed under the Sound category (cs.SD) and is accompanied by links to community-built tools such as the Bibliographic Explorer and Connected Papers, which help readers navigate citation networks [4][5]. Automated speaker verification systems remain vulnerable to synthetic speech and voice conversion attacks, and the ASVspoof challenges have become a standard yardstick for countermeasures. The RAT result suggests that architectures that exploit speaker references during training — even when those references are discarded at test time — can outperform single-utterance baselines [1]. The authors note that the reference channel’s contribution is minimized so thoroughly during optimization that inference becomes “largely independent of the reference channel” [1].