Privacy-Preserving Speaker Verification and Speech Recognition

This paper proposes an approach to speaker verification and speech recognition in environments that require authentication and privacy protection, while accuracy and data utility must remain high. Our methodology aims at protecting audio files and users' identities through the use of encryption and hashing algorithms, while at the same time providing accurate speaker's identity prediction. In addition, for speech recognition, we introduce a mechanism to anonymize the resulting transcript of the recognized spoken language using the Named Entity Recognition method by removing sensitive entities from the text according to the user's preferences. Furthermore, a privacy-preserving version of the original audio is obtained by performing a text-to-speech translation of the anonymized transcript, which together, the anonymous audio and transcript can be transmitted to third parties or service providers without violating privacy restrictions. The proposed methodology has been validated with a set of experiments on a well-known audio dataset, the Librispeech dataset. A type of Time Delay Neural Networks, ECAPA-TDNN was used for speaker verification, Deep Speech as a type of Recurrent Neural Networks was used for speech recognition, NER for entity recognition, cryptography and hashing for privacy protection. The results demonstrate the validity of our approach to protecting the privacy of user data and biometric information while simultaneously performing data analysis with a high degree of accuracy and similarity with the results obtained with no privacy mechanisms in place, also considering the use of several privacy mechanisms.