【摘要】 隐藏说话者真实身份的技术称为说话者匿名化。为了欺骗自动声纹验证(Automatic Speaker Verification, ASV)系统，说话者匿名化通常通过对原始语音的时间或频谱特性来进行修改，例如通过音高缩放、声道长度归一化或语音转换。然而，匿名语音可以通过对ASV进行再训练来恢复识别出匿名语音的真实身份，例如通过匿名化同一说话者的语音来增强数据实现对匿名语音真实身份的识别。为了评估说话人匿名化的有效性，文中提出了一种注册和测试数据的预恢复方法，该方法将音频特征提取的关键参数作为遍历变量，逐一检查其对后端分类网络特征的适配性，适配性高的参数将予以保留。并对匿名语音的反匿名化进行了研究和比较。实验结果表明，预恢复方法对说话者反匿名化是有效的。此外，还发现测试数据的预恢复比注册数据的预还原表现更好。更多还原

【Abstract】 The technique to hide the real identity of speakers is called speaker anonymization. Aiming at deceiving automatic speaker verification(ASV) systems, speaker anonymization is usually conducted by modifying the temporal or spectral properties of original voices, e.g., by pitch scaling, by vocal tract length normalization(VTLN) or by voice conversion(VC). However, the real identity of anonymized speech can be recovered with a careful re-training of ASVs, e.g., data augmentation by anonymizing voices of the same speaker. In order to evaluate the effectiveness of speaker anonymization, a pre-restoration method for both enrollment and testing data is proposed, investigated and compared for the de-anonymization of anonymized voices. Experimental results show that the prerestoration method is effective to speaker deanonymization. Moreover, it is also found that the pre-restoration for testing data performs better than that for enrollment data.更多还原