Abstract:

Synthetic speech attacks can pose a significant threat to people’s lives. To address the issues of the existing models’ lack of the ability to extract key information from redundant data and the limitations of a single model in using the advantages of multiple detection models， a synthetic speech detection model based on Dual branch with Attention Branch and Squeeze-Excitation Inception （SE-Inc） Branch （Dual-ABIB） was proposed. Firstly， the initial feature maps extracted by Sinc-based Convolutional Neural Network （SincNet） were utilized to train the attention branch of the synthetic speech detection model， and the attention maps were output. Secondly， the attention maps were multiplied and superposed with the original feature maps， and the result was trained as the input for the SE-Inc branch. Finally， classification scores obtained by the two branches were processed through decision-level weighted fusion to achieve synthetic speech detection. Experimental results show that the proposed model achieves a minimum tandem Detection Cost Function （min t-DCF） of 0.033 2 and an Equal Error Rate （EER） of 1.15% on ASVspoof2019 dataset when the number of parameters is 539×10³. Compared with SE-ResABNet （Squeeze-Excitation ResNet Attention Branch Network）， when the number of parameters of the proposed model is only 56% of that of SE-ResABNet， the proposed model has the min t-DCF and EER reduced by 34.5% and 39.2% respectively. At the same time， the proposed model shows better generalization ability on ASVspoof2015 and ASVspoof2021 datasets. The above results verify that Dual-ABIB can obtain lower min t-DCF and EER with less of parameters.

Key words: attention mechanism, Squeeze-Excitation (SE) module, dual branch, synthetic speech detection, decision-level fusion

摘要：

合成语音攻击给人们的生活带来巨大的威胁。为了解决现有模型从冗余信息中提取关键信息能力不足和单一模型无法综合利用多检测模型优势的问题，提出一种基于注意力和挤压-激励（SE）模块Inception （SE-Inc）的双分支（Dual-ABIB）合成语音检测模型。首先，基于SincNet（Sinc-based convolutional neural Network）提取的初始特征图训练注意力分支合成语音检测模型，并输出注意力图；其次，将注意力图和初始特征图相乘后再叠加，并将结果作为SE-Inc分支的输入进行训练；最后，通过决策级加权融合处理2个分支获得的分类分数，从而实现合成语音检测。实验结果表明，所提模型在参数量为539×10³的情况下，在ASVspoof2019数据集上获得了0.033 2的最小串联检测代价函数（min t-DCF）和1.15%的等错误率（EER）；与SE-ResABNet （Squeeze-Excitation ResNet Attention Branch Network）相比，所提模型在参数量仅为它的56%的情况下，min t-DCF和EER分别下降了34.5%和39.2%；同时，在ASVspoof2015和ASVspoof2021数据集上所提模型表现了更好的泛化能力。以上结果验证了所提模型能够在参数量较小的情况下，获得更低的min t-DCF和EER。

关键词: 注意力机制, 挤压-激励模块, 双分支, 合成语音检测, 决策级融合