基于上下文信息与注意力特征的欺骗语音检测

Spoof speech detection based on context information and attention feature

随着语音合成和语音转换技术的快速发展,欺骗语音检测方法仍存在欺骗检测准确率低、通用性差等问题。因此,提出一种基于上下文信息与注意力特征的端到端的欺骗检测方法。该方法基于深度残差收缩网络(DRSN),利用双分支上下文信息协调融合模块(DCCM)聚集丰富的上下文信息,融合基于协调时频注意力机制(CTFA)的特征以获得具有上下文信息的跨维度交互特征,从而最大化捕获伪影的潜力。与最佳基线系统相比,在ASVspoof 2019 LA数据集中,所提方法在EER和t-DCF性能指标上分别降低68%和65%;在ASVspoof 2021 LA数据集中,所提方法的EER和t-DCF分别为4.81和0.3115,分别降低48%和10%。实验结果表明,所提方法能有效提高欺骗语音检测的准确率和泛化能力。

With the rapid development of speech synthesis and speech conversion technology,methods of spoof speech detection still have problems such as low spoof detection accuracy and poor generality.Therefore,an end-to-end spoof detection method based on context information and attention feature was proposed.Based on deep residual shrinkage network(DRSN),the proposed method used the dual-branch context information coordination fusion module(DCCM)to aggregate rich context information,and fused features based on coordinate time-frequency attention(CTFA)to obtain cross-dimensional interaction features with context information,thus maximizing the potential of capturing artifacts.Compared with the best baseline system,in the ASVspoof 2019 LA dataset,the proposed method had reduced the EER and t-DCF performance indicators by 68%and 65%respectively,in the ASVspoof 2021 LA dataset,the EER and t-DCF of the proposed method were 4.81 and 0.3115 and dropped by 48%and 10%separately.The experimental results show that this method can effectively improve the accuracy and generalization ability of spoof speech detection.