Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

XLSR-Mamba: A Dual-Column Bidirectional State Space Model for Spoofing Attack Detection

Yang Xiao, Rohan Kumar Das

TL;DR

Spoofing attacks threaten ASV systems, and long-range temporal cues are challenging for Transformer-based models. The paper introduces DuaBiMamba, a dual-column bidirectional state-space module, and XLSR-Mamba, which fuses wav2vec 2.0 XLS-R features with BiMamba blocks to efficiently model long sequences. Evaluations on ASVspoof 2021 LA/DF and In-the-Wild show competitive to state-of-the-art results with superior real-time inference and robust performance in real-world data. Together, these contributions demonstrate that Mamba-based architectures can outperform Transformers in anti-spoofing while enabling faster, scalable deployment.

Abstract

Transformers and their variants have achieved great success in speech processing. However, their multi-head self-attention mechanism is computationally expensive. Therefore, one novel selective state space model, Mamba, has been proposed as an alternative. Building on its success in automatic speech recognition, we apply Mamba for spoofing attack detection. Mamba is well-suited for this task as it can capture the artifacts in spoofed speech signals by handling long-length sequences. However, Mamba's performance may suffer when it is trained with limited labeled data. To mitigate this, we propose combining a new structure of Mamba based on a dual-column architecture with self-supervised learning, using the pre-trained wav2vec 2.0 model. The experiments show that our proposed approach achieves competitive results and faster inference on the ASVspoof 2021 LA and DF datasets, and on the more challenging In-the-Wild dataset, it emerges as the strongest candidate for spoofing attack detection. The code has been publicly released in https://github.com/swagshaw/XLSR-Mamba.

XLSR-Mamba: A Dual-Column Bidirectional State Space Model for Spoofing Attack Detection

TL;DR

Spoofing attacks threaten ASV systems, and long-range temporal cues are challenging for Transformer-based models. The paper introduces DuaBiMamba, a dual-column bidirectional state-space module, and XLSR-Mamba, which fuses wav2vec 2.0 XLS-R features with BiMamba blocks to efficiently model long sequences. Evaluations on ASVspoof 2021 LA/DF and In-the-Wild show competitive to state-of-the-art results with superior real-time inference and robust performance in real-world data. Together, these contributions demonstrate that Mamba-based architectures can outperform Transformers in anti-spoofing while enabling faster, scalable deployment.

Abstract

Transformers and their variants have achieved great success in speech processing. However, their multi-head self-attention mechanism is computationally expensive. Therefore, one novel selective state space model, Mamba, has been proposed as an alternative. Building on its success in automatic speech recognition, we apply Mamba for spoofing attack detection. Mamba is well-suited for this task as it can capture the artifacts in spoofed speech signals by handling long-length sequences. However, Mamba's performance may suffer when it is trained with limited labeled data. To mitigate this, we propose combining a new structure of Mamba based on a dual-column architecture with self-supervised learning, using the pre-trained wav2vec 2.0 model. The experiments show that our proposed approach achieves competitive results and faster inference on the ASVspoof 2021 LA and DF datasets, and on the more challenging In-the-Wild dataset, it emerges as the strongest candidate for spoofing attack detection. The code has been publicly released in https://github.com/swagshaw/XLSR-Mamba.

Paper Structure

This paper contains 15 sections, 2 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Proposed Method
  4. Bidirectional Mamba
  5. Proposed Dual-Column Bidirectional Mamba
  6. Proposed XLSR-Mamba
  7. Experimental Setup
  8. Datasets and Performance Metrics
  9. Implementation Details
  10. Results and Analysis
  11. DuaBiMamba vs. ExtBiMamba & InnBiMamba
  12. XLSR-Mamba vs. Other SOTA Models
  13. Efficiency and Effectiveness Analysis of XLSR-Mamba
  14. Evaluation on In-the-Wild dataset
  15. Conclusion

Figures (3)

  • Figure 1: Overview of the XLSR-Mamba architecture for anti-spoofing, with three different BiMamba configurations: (a) External Bidirectional Mamba (ExtBiMamba); (b) Inner Bidirectional Mamba (InnBiMamba); and (c) Dual-Column Bidirectional Mamba (DuaBiMamba); (d) XLSR-Mamba model pipeline, including XLS-R feature extraction, linear projection, BiMamba blocks, and prediction head. The prediction head is implemented as a linear layer that maps features produced by the BiMamba blocks into whether the input speech is bonafide or spoofed.
  • Figure 2: Inference speed comparison (real-time factor) between XLSR-Mamba and XLSR-Conformer asvconformer across utterance durations from 2 to 10 seconds.
  • Figure 3: The clustering of In-the-Wild test set samples is visualized in 2D t-SNE plots from the model's higher layers.