Project overview
We investigate why some voices are easy or difficult to anonymize, and how this contributes to attitudes of trust toward voice anonymization systems. The mechanisms at work between algorithmic measurements of anonymity (objective metrics) and human perception of anonymity (subjective metrics) are not well understood. Different adversarial scenarios may require optimizing for objective metrics (e.g., attacks on automatic speaker verification systems) while other scenarios require optimizing for subjective metrics (e.g., human eavesdroppers). The problem of achieving trustworthy anonymization capabilities becomes especially difficult because it is not clear what the ground truth should be.
