Research on speech replay attack detection using approach of multi-feature and multi-classifier fusion
Ziheng Cheng, Ye Jiang · 2020
Replay attack detection is crucial to prevent security and reliability when the automatic speaker verification (ASV) systems encounter the spoofed attacks. In order to overcome the shortcomings of traditional cepstrum features can not effectively represent the difference between real speech and replay speech in low-frequency band. We proposed a novel low frequency cepstrum coefficient under singular value decomposition (LCSVD). In addition, an effective feature fusion method (LCSVD+CQCC) based GSV and RELIEF algorithm in the super vector space is studied to improve the performance of detection system. We also proposed to use an multi-classifier set, which includes Random Forest (RF) and LogitBoost classifiers. Experimental results have shown that the proposed multi-feature and multi-classifier fusion system can provide substantially better performance than baseline. EER of primary system on development and evaluation set are respectively 2.49% and 16.58%, compared to baseline 11.28% and 31.84%, which are also better than baseline 77.9% and 47.9%.