摘要
在现实应用场景中,攻击者在伪造语音中加入加性噪声或者混响等干扰,会导致经纯净语音训练得到的检测系统性能急剧下降,为此,通过设计一种激活函数代替残差网络中的跳跃连接,实现了具有噪声鲁棒性的合成语音检测系统。通过分析不同激活函数对残差块跳跃连接的影响,将输入特征划分为非显著特征、显著特征和无法判断特征,提出了一个新的激活函数,并通过方差增长的算法来寻找激活函数的最优参数。实验结果表明,与现有方法相比,不仅显著降低了系统的等错误率,而且对噪声干扰具有很好的鲁棒性。
In real-world application scenarios,attackers often add additive noise or reverberation and other interferences to the spoofing speech,which will cause the performance of the detection system trained with clean speech to drop sharply.Therefore,an activation function was designed to replace the skip connection in the residual network,thereby proposing a synthetic speech detection system with noise-robust.After analyzing the influence of different activation functions on the skip connection of the residual block,the input features were divided into non-significant features,significant features and undetermined features,and a novel activation function was proposed.The optimal parameters of the activation function were determined through the method of variance growth.Experimental results show that compared with existing methods,the method proposed not only significantly reduces the equal error rate of the system,but also has good robustness to noise interference.
作者
杨曼
简志华
梁承涵
Yang Man;Jian Zhihua;Liang Chenghan(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou 310018,China)
出处
《电信科学》
北大核心
2026年第3期70-80,共11页
Telecommunications Science
基金
国家自然科学基金资助项目(No.61201301,No.61772166)。
关键词
伪造语音检测
合成语音检测
激活函数
噪声鲁棒性
spoofing speech detection
synthetic speech detection
activation function
noise-robust
