Evaluating X-Vector-Based Speaker Anonymization Under White-Box Assessment

Abstract

In the scenario of the Voice Privacy challenge, anonymization is achieved by converting all utterances from a source speaker to match the same target identity; this identity being randomly selected. In this context, an attacker with maximum knowledge about the anonymization system can not infer the target identity. This article proposed to constrain the target selection to a specific identity, i.e., removing the random selection of identity, to evaluate the extreme threat under a whitebox assessment (the attacker has complete knowledge about the system). Targeting a unique identity also allows us to investigate whether some target's identities are better than others to anonymize a given speaker.

Anonymizing Speaker Voices: Easy to Imitate, Difficult to Recognize?

“…The complexity is compounded by how speech synthesis is typically evaluated, often relying on human judgements which naturally vary from listener to listener and are difficult to replicate over decades of research [3,17,18]. Voice anonymization can also be framed as a special type of voice conversion task [6] -one where the original speaker shall not be revealed. Still, voice privacy is evaluated along different metrics or based on different assumptions than voice conversion.…”

“…The system D [27] is derived from baseline B2. Systems based on x-vectors include: A [28], M [29], O [30], and S [31]. While system I 2 uses modifications to formants, system K 3 combines x-vectors, speaker similarity models, and a voice indistinguishability metric.…”

“…A prominent open question in speech technology is understanding how individual speakers impact downstream tasks, spanning fields such as text-to-speech (TTS) synthesis [1], voice conversion [2], speech synthesis evaluation [3], automatic speaker verification (ASV) [4], automatic speech recognition (ASR) [5], and voice anonymization [6,7]. Anonymization techniques are primarily evaluated based on their performance in ASV [8].…”

Vocoder drift compensation by x-vector alignment in speaker anonymisation

“…For the majority of anonymisation systems proposed to date, this requirement is fulfilled by maximising some measure of the distance between the chosen pseudo-speaker embedding and the original speaker embedding. The most popular anonymisation function to date uses a pool of external x-vectors [2,3,4,5,6,7]. The pseudo-speaker embedding is obtained by averaging a random subset of the furthest x-vectors in the pool from the x-vector of the original speaker.…”