Device feature, which contains the information of both recording channel and playback channel, is the critical trait for replay spoofing detection. So far there have not been any technical reports about the usage of device information in spoofing detection for speaker verification. In this paper, we propose to build a replay device feature (RDF) extractor on the basis of the genuine-replay-pair training database. The RDF extractor is trained in constant-Q transform (CQT) spectrum domain. A bidirectional long short-term memory (BLSTM) is used in the neural network and finally the RDF extractor is formed by applying discrete cosine transform (DCT) to the output vector of the BLSTM. The experimental result on ASVspoof 2017 corpus version 2.0 shows that equal error rate (EER) of replay detection system with the proposed RDF reaches 15.08%. Furthermore, by combining the RDF with constant-Q cepstral coefficients plus log energy (CQCCE), the EER of the detection system can reduce to 8.99%. In addition, the experimental results also show that the RDF has much complementarity with conventional features.
Cite as: You, C.H., Yang, J., Tran, H.D. (2019) Device Feature Extractor for Replay Spoofing Detection. Proc. Interspeech 2019, 2933-2937, doi: 10.21437/Interspeech.2019-2137
@inproceedings{you19c_interspeech,
author={Chang Huai You and Jichen Yang and Huy Dat Tran},
title={{Device Feature Extractor for Replay Spoofing Detection}},
year=2019,
booktitle={Proc. Interspeech 2019},
pages={2933--2937},
doi={10.21437/Interspeech.2019-2137}
}