Theoretical vulnerabilities in map speaker adaptation
Tetsushi Ohki, Akira Otsuka · 2017
We analyze the theoretical vulnerability of maximum a posteriori(MAP) speaker adaptation, which is widely used in practical speaker recognition systems. First, we proved that there exist a set of feature vectors, what are called wolves, which can impersonate almost all the registered speakers with probability asymptotically close to 1 with at most two trials. Second, our experiment shows that the wolves with appropriate parameters achieved 0.99 of successful impersonation rate on Spear speaker recognition toolkit with ATR speech database.