ATMM-SAGA: Alternating Training for Multi-Module with Score-Aware Gated Attention SASV system

TL;DR

This paper introduces ATMM-SAGA, a novel SASV system that uses score-aware gated attention to effectively combine speaker verification and anti-spoofing scores, significantly improving robustness against spoofing attacks.

Contribution

It proposes a new score-aware gated attention fusion scheme for SASV, enhancing the integration of countermeasure scores with speaker embeddings.

Findings

Achieves low SASV-EER of around 2.2% on ASVspoof2019 dataset.

Demonstrates improved anti-spoofing robustness over baseline methods.

Validates effectiveness of score-aware gating in SASV systems.

Abstract

The objective of automatic speaker verification (ASV) systems is to determine whether a given test speech utterance corresponds to a claimed enrolled speaker. These systems have a wide range of applications, and ensuring their reliability is crucial. In this paper, we propose a spoofing-robust automatic speaker verification (SASV) system employing a score-aware gated attention (SAGA) fusion scheme, integrating scores from a pre-trained countermeasure (CM) with speaker embeddings from a pre-trained ASV. Specifically, we employ the AASIST and ECAPA-TDNN models. SAGA acts as an adaptive gating mechanism, where the CM score determines how strongly ASV embeddings influence the final SASV decision. Experiments on the ASVspoof2019 logical access dataset demonstrate that the proposed SASV system achieves an SASV equal error rate (SASV-EER) and agnostic detection cost function (a-DCF) of 2.31%, 0.0603 for the development set and 2.18%, 0.0480 for the evaluation set.