Replay Attack Detection in Speaker Verification Using non-voiced segments and Decision Level Feature Switching
Saranya M S, R. Padmanabhan, Hema A. Murthy · 2018
This paper proposes a novel approach for replay attack detection, using reverberation and channel information from non-voiced (silence and unvoiced) segments of utterances. The non-voiced segments are determined using a voice activity detector. These non-voiced segments are likely to contain reverberation and channel information. Multiple feature representations are used to capture the remnant vocal tract information in non-voiced segments. Gaussian mixture models are used to build three different baseline systems corresponding to that of three different features. Voting is performed to decide whether a given input utterance is replayed or not. Equal error rate (EER) is computed using the likelihood ratio of the genuine and spoofed model from the best baseline system. Evaluation on the ASV-Spoof-2017 challenge dataset shows that the proposed approach outperforms the best baseline system with a relative improvement of 37% in terms of EER.