Discriminative Feature Extraction Based on SWM for Playback Attack Detection

This paper proposes discriminative features for playback attack detection (PAD) based on statisticallyweighted magnitude spectrum (SWM). The proposed SWM method aims to magnify differences between playback and genuine speech through a discriminative feature, namely, statistically weighted magnitude spectrum constant-Q cepstrum coefficient (SWM-CQCC). The ASVspoof 2017 evaluation set is used to evaluate the proposed method. Our experimental results show that SWM-CQCC can outperforms most of all existing methods we are aware of including published implementations of I-Vector and VESAIFCC.