RAT: Reference-Augmented Training for ASV Anti-Spoofing

Authors: Vojtěch Staněk, Anton Firc, Jakub Reš, Kamil Malinka

Published: 2026-06-09 14:20:05+00:00

Comment: Accepted to Interspeech 2026

AI Summary

This paper introduces Reference-Augmented Training (RAT) for ASV anti-spoofing, which conditions a countermeasure architecture on speaker-reference recordings during training. Surprisingly, the model converges to a solution that largely ignores the reference during inference, yet training with it induces an invariance that significantly improves deepfake detection. RAT achieves state-of-the-art performance on the ASVspoof 5 benchmark, surpassing large ensemble systems, even when the reference is absent or degraded at inference.

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.

Key findings

RAT achieves state-of-the-art performance on ASVspoof 5 with a single detector, obtaining 2.57% EER and 0.074 minDCF, outperforming larger ensemble systems. A crucial finding is that the model's performance remains robust and high even when the reference recording is replaced with a zero vector, noise, or a mismatched speaker during inference. Analysis showed that the reference's contribution rapidly diminishes during training, leading to inference largely independent of the reference channel.

Approach

They propose a Reference-Augmented Training (RAT) strategy utilizing a Reference-Informed Block (RIB) architecture. This block takes both test and reference speech, extracting features with a shared XLS-R model, and processes them through parallel MLP and Multi-Head Cross-Attention branches. The core insight is that while the reference is used during training, the optimization process diminishes its contribution, leading to a detector that effectively operates as a single-utterance system during inference, maintaining high performance.

Datasets

ASVspoof 5

Model(s)

XLS-R (300M parameters) as a feature extractor, Reference-Informed Block (RIB) comprising an MLP branch and a Multi-Head Cross-Attention branch, and a 3-layer MLP classifier.

Author countries

Czech Republic

← Previous