基于时延神经网络的语音欺骗检测模型
Voice spoofing detection model based on time-delay neural network
刘景贤 1李丽香 1申晋瑜1
作者信息
- 1. 北京邮电大学网络与交换技术国家重点实验室信息安全中心，北京，100876
- 折叠
摘要
随着语音合成与转换技术的迅猛发展，自动说话人验证系统面临着日益严峻的语音欺骗攻击威胁。然而，现有检测模型在复杂声学环境（如背景噪声、信道失真）下普遍存在鲁棒性下降与判别能力不足的瓶颈，难以有效捕捉欺骗语音中微弱且多变的声学线索。针对上述问题，本文提出了一种上下文感知注意力融合时延神经网络。该方法通过三级SE-Res2Block模块提取多尺度时序特征，设计上下文感知注意力融合模块实现关键上下文表征的动态强化，并采用向下融合策略整合深浅层特征。实验结果表明，所提模型在 ASVspoof2019 LA数据集上显著优于基线系统，尤其在未知攻击类型及复杂噪声条件下，等错误率降至 4.28%，验证了所提方法的有效性与泛化能力。
Abstract
With the rapid advancement of voice synthesis and conversion technologies, automatic speaker verification systems face increasingly severe threats from voice spoofing attacks. However, existing detection models generally suffer from bottlenecks of reduced robustness and insufficient discriminative capability in complex acoustic environments (e.g., background noise, channel distortion), making it difficult to effectively capture the subtle and variable acoustic cues in spoofed speech. To address these issues, this paper proposes a Context-Aware Attention Fusion Time Delay Neural Network. This method extracts multi-scale temporal features using a three-level SE-Res2Block module and designs a Context-Aware Attention Fusion Module to dynamically strengthen key contextual representations. Simultaneously, a downward fusion strategy is introduced to integrate deep and shallow features, enhancing the model\'s perception of subtle forgery traces. Experimental results demonstrate that our model significantly outperforms baseline systems on the ASVspoof2019 LA dataset. Particularly under conditions of unknown attack types and complex noise, the Equal Error Rate drops to 4.28%, verifying the effectiveness and generalization capability of the proposed method.关键词
语音欺骗检测/时延神经网络/注意力机制/多尺度特征融合Key words
Voice Spoofing Detection/Time Delay Neural Network/Attention Mechanism/Multi - scale Feature Fusion引用本文复制引用
刘景贤,李丽香,申晋瑜.基于时延神经网络的语音欺骗检测模型[EB/OL].(2026-04-08)[2026-04-10].http://www.paper.edu.cn/releasepaper/content/202604-73.学科分类
计算技术、计算机技术
评论