Are disentangled representations all you need to build speaker anonymization systems?

Abstract

Speech signals contain a lot of sensitive information, such as the speaker's\nidentity, which raises privacy concerns when speech data get collected. Speaker\nanonymization aims to transform a speech signal to remove the source speaker's\nidentity while leaving the spoken content unchanged. Current methods perform\nthe transformation by relying on content/speaker disentanglement and voice\nconversion. Usually, an acoustic model from an automatic speech recognition\nsystem extracts the content representation while an x-vector system extracts\nthe speaker representation. Prior work has shown that the extracted features\nare not perfectly disentangled. This paper tackles how to improve features\ndisentanglement, and thus the converted anonymized speech. We propose enhancing\nthe disentanglement by removing speaker information from the acoustic model\nusing vector quantization. Evaluation done using the VoicePrivacy 2022 toolkit\nshowed that vector quantization helps conceal the original speaker identity\nwhile maintaining utility for speech recognition.\n

Topics & Concepts