Differentially Private Speaker Anonymization

Sharing real-world speech utterances is key to the training and deployment of voice-based services. However, it also raises privacy risks as contains a wealth personal data. Speaker anonymization aims remove speaker information from utterance while leaving its linguistic prosodic attributes intact. State-of-the-art techniques operate by disentangling (represented via embedding) these re-synthesizing based on embedding another speaker. Prior research in community has shown that often provides brittle protection, even less so any provable guarantee. In this work, we show disentanglement indeed not perfect: still contain information. We introducing differentially private feature extractors an autoencoder automatic recognizer, respectively, trained using noise layers. plug state-of-the-art pipeline generate, for first time, with upper bound they contain. evaluate empirically utility resulting our approach LibriSpeech data set. Experimental results generated retain very high recognition inference, being much better protected against strong adversaries who leverage full knowledge process try infer identity.