Improving Short Utterance Anti-Spoofing with AASIST2

TL;DR

The paper tackles spoof speech detection under short utterance durations, where duration mismatch degrades performance. It introduces AASIST2, which replaces ResNet blocks with Res2Net blocks to capture multi-scale temporal features and uses wav2vec 2.0 XLS-R with AM-Softmax for stronger discrimination. To further handle duration variability, the authors formulate Dynamic Chunk Size (DCS) and Adaptive Large Margin Fine-Tuning (ALMFT), implementing Margin = $A \times Duration + B$ with $A=\frac{3}{50}$ and $B=\frac{7}{50}$. Experiments on ASVspoof 2019 LA/2021 LA/DF show that AASIST2 achieves SOTA performance on 2019 LA and improves short-utterance evaluation across datasets, while maintaining standard performance on longer utterances. The approach offers practical improvements for real-world deployment with varied speech lengths.

Abstract

The wav2vec 2.0 and integrated spectro-temporal graph attention network (AASIST) based countermeasure achieves great performance in speech anti-spoofing. However, current spoof speech detection systems have fixed training and evaluation durations, while the performance degrades significantly during short utterance evaluation. To solve this problem, AASIST can be improved to AASIST2 by modifying the residual blocks to Res2Net blocks. The modified Res2Net blocks can extract multi-scale features and improve the detection performance for speech of different durations, thus improving the short utterance evaluation performance. On the other hand, adaptive large margin fine-tuning (ALMFT) has achieved performance improvement in short utterance speaker verification. Therefore, we apply Dynamic Chunk Size (DCS) and ALMFT training strategies in speech anti-spoofing to further improve the performance of short utterance evaluation. Experiments demonstrate that the proposed AASIST2 improves the performance of short utterance evaluation while maintaining the performance of regular evaluation on different datasets.

Figures (2)

• Figure 1: The illustration of the residual block in AASIST and Res2Net block in AASIST2.
• Figure 2: The distribution of speech duration in the datasets.