Scholarlink

Spoofing detection models suffer significant accuracy degradation on unknown attacks, especially in cross-dataset scenarios. To address this issue, we propose a MoE-based self-supervised spoofing speech detection model, WavZip, which leverages Wav2vec2 for deep feature extraction, incorporates the squeeze-and-excitation mechanism into Zipformer blocks, and extends the original Zipformer to multiple expert models. By introducing a gating mechanism and load balancing loss, the model adaptively allocates expert weights based on input characteristics, improving collaboration efficiency and optimizing workload distribution. Additionally, to enhance robustness, data augmentation was performed by intra-class component shuffling. Cross-dataset evaluation results show that the proposed model achieves significantly lower equal error rate (EER) and minimum tandem detection cost function (min t-DCF) than the state-of-the-art (SOTA) model on both the ASVspoof2021 and In-the-Wild (ITW) datasets. Specifically, in the ASVspoof2021 LA scenario, the model shows a 90% improvement over RawNet2 and a 36% improvement over the XLSR-Conformer model.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhanced Spoofing Speech Detection with Self-supervised Feature Extraction and Multi-expert Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yanhong Long,
  • Guangcun Wei,
  • Chengde Zhang,
  • Boyan Guo,
  • Wenjing Wang,
  • Guanghao Liu

摘要

Spoofing detection models suffer significant accuracy degradation on unknown attacks, especially in cross-dataset scenarios. To address this issue, we propose a MoE-based self-supervised spoofing speech detection model, WavZip, which leverages Wav2vec2 for deep feature extraction, incorporates the squeeze-and-excitation mechanism into Zipformer blocks, and extends the original Zipformer to multiple expert models. By introducing a gating mechanism and load balancing loss, the model adaptively allocates expert weights based on input characteristics, improving collaboration efficiency and optimizing workload distribution. Additionally, to enhance robustness, data augmentation was performed by intra-class component shuffling. Cross-dataset evaluation results show that the proposed model achieves significantly lower equal error rate (EER) and minimum tandem detection cost function (min t-DCF) than the state-of-the-art (SOTA) model on both the ASVspoof2021 and In-the-Wild (ITW) datasets. Specifically, in the ASVspoof2021 LA scenario, the model shows a 90% improvement over RawNet2 and a 36% improvement over the XLSR-Conformer model.