Speaker anonymization using neural audio codec language models

TL;DR

This paper introduces a novel speaker anonymization method using neural audio codecs and language models, aiming to better obfuscate speaker identity by leveraging quantized codes that bottleneck speaker information.

Contribution

It proposes a NAC-based speaker anonymization approach that effectively reduces speaker information in speech signals, addressing limitations of previous methods involving speaker embedding perturbation.

Findings

NAC-based anonymization reduces speaker identity leakage.

The approach achieves high-quality synthetic speech with anonymization.

Evaluation shows promising results on Voice Privacy Challenge 2022.

Abstract

The vast majority of approaches to speaker anonymization involve the extraction of fundamental frequency estimates, linguistic features and a speaker embedding which is perturbed to obfuscate the speaker identity before an anonymized speech waveform is resynthesized using a vocoder. Recent work has shown that x-vector transformations are difficult to control consistently: other sources of speaker information contained within fundamental frequency and linguistic features are re-entangled upon vocoding, meaning that anonymized speech signals still contain speaker information. We propose an approach based upon neural audio codecs (NACs), which are known to generate high-quality synthetic speech when combined with language models. NACs use quantized codes, which are known to effectively bottleneck speaker-related information: we demonstrate the potential of speaker anonymization systems based on NAC language modeling by applying the evaluation framework of the Voice Privacy Challenge 2022.