基于Transformer编码器的合成语声检测系统

Transformer encoder-based spoofing countermeasure for synthetic speechdetection

自动说话人认证系统是一种常用的目标说话人身份认证方案,但它在合成语声的攻击下表现出脆弱性,合成语声检测系统试图解决这一问题。该文提出了一种基于Transformer编码器的合成语声检测方法,利用自注意力机制学习输入特征内部的长期依赖关系。合成语声检测问题并不关注句子的抽象语义特征,用参数量较小的模型也能得到较好的检测性能。该文分别测试了4种常用合成语声检测特征在Transformer编码器上的表现,在国际标准的ASVspoof2019挑战赛的逻辑攻击数据集上,基于线性频率倒谱系数特征和Transformer编码器的系统等错误率与串联检测代价函数分别为3.13%和0.0708,且模型参数量仅为0.082 M,在较小参数量下得到了较好的检测性能。

The automatic speaker verification system is a commonly used solution for target speaker identity authentication, but it shows vulnerability under the attack of synthetic speech, which can be alleviated by a spoofing countermeasure system. In this paper, we introduce a synthetic speech detection method based on the Transformer encoder, which uses the self-attention mechanism to learn the long-term dependencies of the input features. Synthetic speech detection does not focus on the abstract semantic features of the sentences, and a model with small parameters can also perform well. This paper evaluated the performance of four commonly used synthetic speech detection features on Transformer encoders. On the evaluation set of the ASVspoof2019challenge logical access scenario, the proposed system based on linear frequency cepstral coefficient features and Transformer encoder achieves an equal error rate(EER) of 3.13% and a tandem detection cost function(tDCF) of 0.0708, respectively, and the parameters of the model is only 0.082 M, a better detection performance is obtained with a smaller model.