Replay Attack Detection Using Generalized Cross-Correlation of Stereo Signal

In this paper, we propose a replay attack detection method using the generalized cross-correlation (GCC) of a stereo signal for automatic speaker verification. In particular, this method focuses on a specific replay attack characteristics when speech is not active. In a genuine speaker case, when speech is not active, the maximum value of GCC is low since surrounding noise arrives from any direction. In contrast, in a replay attack case, even when the played speech is not active, the maximum value of GCC is high since recorded noise or electromagnetic noise is played by a loudspeaker for replay attack. Based on this assumption, two approaches of replay attack detection are introduced. One is to use the minimum value of GCC in short pauses. The other one is to use the average value of GCC in silent periods before the start point and after the end point of a target utterance. In experiments, it is confirmed that the proposed methods achieve low error rates without environmental restrictions.