Improving the Speaker Anonymization Evaluation's Robustness to Target Speakers with Adversarial Learning
This work addresses a specific issue in privacy evaluation for speaker anonymization, making it incremental but important for improving robustness in speech processing applications.
The paper tackled the problem of overestimated privacy in speaker anonymization evaluation by proposing a target classifier to measure target speaker influence, which can be removed with adversarial learning. Experiments showed this approach is effective for multiple anonymizers, particularly with same-gender target selection, leading to more reliable assessment.
The current privacy evaluation for speaker anonymization often overestimates privacy when a same-gender target selection algorithm (TSA) is used, although this TSA leaks the speaker's gender and should hence be more vulnerable. We hypothesize that this occurs because the evaluation does not account for the fact that anonymized speech contains information from both the source and target speakers. To address this, we propose to add a target classifier that measures the influence of target speaker information in the evaluation, which can also be removed with adversarial learning. Experiments demonstrate that this approach is effective for multiple anonymizers, particularly when using a same-gender TSA, leading to a more reliable assessment.