Multiple Phase Information Combination for Replay Attacks Detection

Abstract

In recent years, the performance of Automatic Speaker Verification (ASV) systems has been improved significantly. However, they are still affected by different kind of spoofing attacks. In this paper, we propose a method that fused different phase features and amplitude features to detect replay attacks. We apply the mel-scale relative phase feature and source-filter vocal tract feature in phase domain for replay attacks detection. These two phase-based features are combined to get complementary information. In addition to these phase characteristics, constant Q cepstral coefficients (CQCCs) are used. The proposed methods are evaluated using the ASVspoof 2017 challenge database, and Gaussian mixture model was used as the back-end model. The proposed approach achieved 55.6% relative error reduction rate than the conventional magnitude-based feature.

Replay attack detection with auditory filter-based relative phase features

“…In the present study, we extend our previous paper [31] by performing more experiments and analysis. The contributions of this study are as follows: (1) According to [31], the detection model in the previous work, which used only a training subset of the ASVspoof 2017 database, might not lead to an impressive result on the evaluation of subset-based testing; hence, we incorporate a training subset and development subset to train the model in our system. (2) In addition to the Mel-filter bank represented to scale the RP information in the previous work, we also apply a gammatone filter bank, thereby simulating the auditory system of humans to convert the important information of the full-band RP-based feature in the gammatone scale, where the RP feature with the reduced dimension is called the gammatone-scale RP.…”

“…Additionally, preliminary experiments have indicated that the Mel-scale RP provides better performance compared with the MGD cepstral coefficient (MGDCC) and MFCC, and CQCC. However, the detailed Mel-scale RP information extraction and analysis were not described in [31]. In the present study, we extend our previous paper [31] by performing more experiments and analysis.…”

“…However, the detailed Mel-scale RP information extraction and analysis were not described in [31]. In the present study, we extend our previous paper [31] by performing more experiments and analysis. The contributions of this study are as follows: (1) According to [31], the detection model in the previous work, which used only a training subset of the ASVspoof 2017 database, might not lead to an impressive result on the evaluation of subset-based testing; hence, we incorporate a training subset and development subset to train the model in our system.…”

Discriminative features based on modified log magnitude spectrum for playback speech detection

“…• Spectral centroid-based features: which include subband spectral centroid frequency coefficients and subband spectral centroid magnitude coefficients [12] and spectral centroid deviation [16]. • Phased-based features: which include instantaneous frequency cosine coefficient [44,45] and modified group delay cepstral coefficient [15]. • Zero time windowing-based features: zero time windowing cepstral coefficients [46,47].…”

“…Since the ASVspoof 2017 challenge [1,2], more and more researchers begin to focus on playback speech detection [3][4][5][6][7][8][9][10]. Similar to many speech signal processing systems, most of all playback speech detection systems usually consist of front-end feature and back-end classifier [11][12][13][14][15][16][17][18]. For the end-to-end systems such as *Correspondence: xlt@dhu.edu.cn † Jichen Yang and Longting Xu contributed equally to this work.…”