RAT: Reference-Augmented Training for ASV Anti-Spoofing

2026-06-09

AI Summary

This work addresses the vulnerability of automatic speaker verification (ASV) systems to deepfake speech attacks, where existing single-sentence detection methods exhibit limited performance. The authors propose a Reference-Augmented Training (RAT) strategy that incorporates speaker reference utterances as conditional inputs during training. By employing a reference-conditioned neural architecture with end-to-end optimization, the model dynamically downweights the contribution of the reference channel. Notably, although the reference signal is discarded during inference, the model implicitly learns robust, spoof-invariant feature representations, revealing a mechanism by which reference signals induce invariance during training. Departing from conventional paradigms that rely on reference signals at test time, the proposed method achieves state-of-the-art performance on the ASVspoof 2021 LA task with a single detector, attaining an EER of 2.57% and a minDCF of 0.074—surpassing existing large ensemble systems.

Abstract

We introduce a spoofing countermeasure architecture conditioned on speaker-reference recordings, but observe that it converges to a solution that effectively ignores the reference during inference. Surprisingly, training with a reference channel induces invariance that improves deepfake detection, even when the reference is absent or mismatched during inference. Based on this observation, we propose a Reference-Augmented Training (RAT) strategy. RAT yields improved detection performance compared to single-utterance baselines, even when the reference recording is replaced with a zero vector at inference. Through rigorous analysis, we demonstrate that the optimization process rapidly diminishes the reference contributions, leading to inference largely independent of the reference channel. Using RAT, we achieve state-of-the-art 2.57% EER and 0.074 minDCF on the ASVspoof 5 benchmark with a single detector, surpassing even large ensemble systems.

Problem

• ASV anti-spoofing
• deepfake detection
• speaker verification
• spoofing countermeasure
• reference-augmented training

Innovation

• Reference-Augmented Training
• anti-spoofing
• deepfake detection
• speaker verification
• reference invariance