Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks

Jee-weon Jung, Hee-Soo Heo, Hemlata Tak, Hye-jin Shim, Joon Son Chung, Bong-Jin Lee, Ha-Jin Yu, Nicholas Evans

TL;DR

AASIST tackles robust audio anti-spoofing in the logical-access regime by unifying spectral and temporal artefact modeling in a single end-to-end graph-based framework. It introduces a heterogeneous stacking graph attention layer (HS-GAL) and a max graph operation (MGO) with a stack-node readout to integrate information across spectro-temporal graphs. The approach yields substantial improvements over prior single-model baselines (≈20% relative min-tDCF) and includes a lightweight variant with 85K parameters that remains competitive. This work enables efficient, non-ensemble spoofing detection suitable for deployment in real-world speaker verification systems.

Abstract

Artefacts that differentiate spoofed from bona-fide utterances can reside in spectral or temporal domains. Their reliable detection usually depends upon computationally demanding ensemble systems where each subsystem is tuned to some specific artefacts. We seek to develop an efficient, single system that can detect a broad range of different spoofing attacks without score-level ensembles. We propose a novel heterogeneous stacking graph attention layer which models artefacts spanning heterogeneous temporal and spectral domains with a heterogeneous attention mechanism and a stack node. With a new max graph operation that involves a competitive mechanism and an extended readout scheme, our approach, named AASIST, outperforms the current state-of-the-art by 20% relative. Even a lightweight variant, AASIST-L, with only 85K parameters, outperforms all competing systems.

AASIST: Audio Anti-Spoofing using Integrated Spectro-Temporal Graph Attention Networks

TL;DR

AASIST tackles robust audio anti-spoofing in the logical-access regime by unifying spectral and temporal artefact modeling in a single end-to-end graph-based framework. It introduces a heterogeneous stacking graph attention layer (HS-GAL) and a max graph operation (MGO) with a stack-node readout to integrate information across spectro-temporal graphs. The approach yields substantial improvements over prior single-model baselines (≈20% relative min-tDCF) and includes a lightweight variant with 85K parameters that remains competitive. This work enables efficient, non-ensemble spoofing detection suitable for deployment in real-world speaker verification systems.

Abstract

Artefacts that differentiate spoofed from bona-fide utterances can reside in spectral or temporal domains. Their reliable detection usually depends upon computationally demanding ensemble systems where each subsystem is tuned to some specific artefacts. We seek to develop an efficient, single system that can detect a broad range of different spoofing attacks without score-level ensembles. We propose a novel heterogeneous stacking graph attention layer which models artefacts spanning heterogeneous temporal and spectral domains with a heterogeneous attention mechanism and a stack node. With a new max graph operation that involves a competitive mechanism and an extended readout scheme, our approach, named AASIST, outperforms the current state-of-the-art by 20% relative. Even a lightweight variant, AASIST-L, with only 85K parameters, outperforms all competing systems.

Paper Structure

This paper contains 14 sections, 1 equation, 1 figure, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. RawNet2-based encoder
  4. Graph module
  5. AASIST
  6. Graph combination
  7. HS-GAL
  8. Max graph operation and readout
  9. Lightweight variant of AASIST
  10. Experiments and results
  11. Dataset and metrics
  12. Implementation details
  13. Results
  14. Conclusion

Figures (1)

  • Figure 1: Overall framework of the proposed AASIST. Identical totak2021end: encoder extracts $F$ and two graph modules each model spectral and temporal domains. Proposed: then, the proposed max graph operation technique adopts two branches that model heterogeneous graphs in parallel, followed by an element-wise maximum. Each branch includes two proposed HS-GAL layers and two graph pooling layers (graph pooling layers and one HS-GAL layer is omitted in the illustration). Finally, the maximum and average of nodes, and the stack node are concatenated followed by an output layer.