声真似が話者照合に与える影響と物真似音声の音響特徴の分析 (音声)

It is quite important to recognize the vulnerability of speaker verification (SV) systems against voice mimicry attacks and to propose countermeasure methods against them. This paper describes an acoustic feature analysis using imitated speech uttered by non-professional speakers and a performance degradation of an HMM-based SV system by the mimicry attacks using the imitated speech. We collected normal/imitated speech uttered by 6 males and 6 females, and analyzed the cepstral feature changes by the imitation. The analysis results show the tendency that the acoustic features of the speakers' voice significantly change by their efforts of the imitation, whereas the distance of the acoustic features between imitated and target speaker's speech is still large. The experimental results on the SV system performance show that the mimicry attacks increase equal error rates by 1.2 to 2.7 times. These facts indicate that non-professional mimicry attacks rarely succeed in spoofing the SV system and yield actual performance degradations in spite of the situation where there is a large gap of acoustic features between imitated and target speech.