基于差分隐私的说话人匿名化方法

he speaker information contained in the voice is an identifiable biological template, which can be used by attackers to steal the unprotected user privacy information. This paper proposes a speaker anonymization method based on the baseline method of the VoicePrivacy Challenge and the privacy metric method in generalized differential privacy to protect the voice privacy while maintaining the voice availability. This method decomposes the target speech, uses differential privacy method to perturb the speaker feature vector, and finally synthesizes the anonymized speech. The experimental results show that the anonymized speech can not only improve the privacy of speech, but also maintain a high level of naturalness and intelligibility, and can complete various downstream tasks in speech.