基于深度学习的说话人匿名化方法

With the increase of privacy legislation, the protection of personal identity information has become increasingly important, especially biometric data such as voice. Speech may contain sensitive information such as gender and identity, posing a risk of data leakage. Users are unwilling to expose this information to untrusted third parties, so privacy protection of voice data is particularly important. This article explores current research methods for voice data privacy protection, such as cryptographic based protection, distributed learning, and anonymization methods. A speaker anonymization scheme based on deep learning is proposed, which combines TF-IDF and TextRank sensitive word extraction algorithms to improve the security of speech content. At the same time, the x-vector model was optimized, and the improved WGAN algorithm was introduced to enhance the anonymity processing capability of voiceprint features. The experimental results show that the method proposed in this paper performs well in both keyword extraction and speaker feature extraction tasks, significantly outperforming traditional methods and having good application prospects.