• Publications
  • Influence
The VoicePrivacy 2022 Challenge Evaluation Plan
TLDR
The voice anonymization task selected for the VoicePrivacy 2020 Challenge is formulated and the datasets used for system development and evaluation are described, including the attack models and the associated objective and subjective evaluation metrics.
A Study of F0 Modification for X-Vector Based Speech Pseudonymization Across Gender
TLDR
It is found that the proposed F0 modification always improves pseudonymization and it is observed that both source and target speaker genders affect the performance gain when modifying the F0.
Speaker information modification in the VoicePrivacy 2020 toolchain
TLDR
This work proposed to replace the triphone-based bottleneck features extractor that requires supervised training by an end-to-end Automatic Speech Recognition (ASR) system, and explored the use of adversarial and semi-adversarial training to learn linguistic features while masking speaker information.
On the Invertibility of a Voice Privacy System Using Embedding Alignment
TLDR
It is shown that a complex system like the baseline of the Voice Privacy Challenge can be approximated by a rotation, estimated using a limited set of $x$-vectors, and the proposed method can recover up to 62% of the speaker identities from anonymized embeddings.
Privacy-Preserving Speech Representation Learning using Vector Quantization
TLDR
This paper proposes to use vector quantization to constrain the representation space and induce the network to suppress the speaker identity, to produce an anonymous representation while preserving speech recognition performance.
Evaluation of Speaker Anonymization on Emotional Speech
TLDR
The results show that the VPC baseline system does not suppress speakers’ emotions against informed attackers, and the emotion recognition performance is degraded by 15% relative to IEMOCAP data, similar to the degradation observed for automatic speech recognition used to evaluate the preservation of the linguistic information.
Evaluating X-Vector-Based Speaker Anonymization Under White-Box Assessment
TLDR
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 white-box assessment (the attacker has complete knowledge about the system).