Abstract
With the increasing use of voice assistants and spoken language interfaces, important concerns have arisen regarding the privacy of voice data. In an attempt to mitigate the threat of attacks on voice data, in this paper we propose an x-vector identification and authentication system. This method modifies the speaker's pitch and accent information from the original speech signal. The proposed method produces a more natural-sounding anonymized voice in addition to the de-identified speaker. However, it yields a relative EER improvement of up to 19.30% for the identified-anonymized registration-test pairs. We observed that along with good speaker de-identification, the anonymized utterances have adequate speech intelligibility and naturalness. Our method can be easily integrated with others as an identification component of a system and eliminates the need to assign speakers to use during identification.