SUT System Description for Anti-Spoofing 2017 Challenge

SUT System Description for Anti-Spoofing 2017 Challenge Department of Computer Engineering Sharif University of Technology, Tehran, Iran Mohammad Adiban, Hossein Sameti, Nooshin Maghsoodi, Sajjad Shahsavari adiban@ce.sharif.edu, sameti@sharif.edu, nmaghsoodi@ce.sharif.edu, mrshahsavari@ce.sharif.edu Abstract Reliability of Automatic Speaker Verification (ASV) systems has always been a concern in dealing with spoofing attacks. Among these attacks, replay attack is the simplest and the easiest accessible method. This paper describes a replay spoofing detection system applied to ASVspoof2017 corpus. To reach this goal, features such as Constant-Q Cepstral Coeffi- cients (CQCC), Modified Group Delay (MGD), Mel Frequency Cepstral Coefficients (MFCC), Relative Spectral Perceptual Linear Predictive (RASTA-PLP) and Linear Prediction Cepstral Coefficients (LPCC), and different classifiers including Gaussian Mixture Models (GMM), Multi- Layer Perceptron (MLP), Support Vector Machine (SVM) and Linear Gaussian (LG) classifier have been employed. We also used identity vector (i-vector) based utterance representation. Finally, scores of different subsystems have been fused to construct the proposed system. The results show that the best performance is attained using this score level fusion. Keywords: Spoofing Attack, Automatic Speaker Verification, Replay, ASVspoof2017 1. Introduction In recent years, there has been growing interest to develop Automatic Speaker Verification sys- tems. ASV systems aim to verify the speaker’s identity using his/her speech. There are many mass-market applications of ASV systems, such as phone banking and trading, password reset- ting, access control in smart phones, credit card activation etc. [1, 2]. Due to the development of ASV systems, concerns about spoofing attacks and security of these systems are increas- ing. To study spoofing approaches and their threats, the ASVspoof2015 [3] and ASVspoof2017 [4] challenges are introduced. Spoofing attacks include four main approaches: impersonation, speech synthesis, voice conversion and replay [3, 5]. Among them, replay attack is the most easily accessible approach available with low technology recording devices, such as voice recorder, laptop, smartphone etc. [5]. These devices may incur different noises during spoofing attacks such as channel or convolutional noise and quantization noise. The channel noise could be originated from the different recording devices, different recording environment and changes in the distance to the microphone. Therefore, in spoofing attacks context, this noise occurs owing to the fact that replayed speech is recorded by two The 2017 Conference on Computational Linguistics and Speech Processing ROCLING 2017, pp. 264-275  The Association for Computational Linguistics and Chinese Language Processing 264 devices and one loudspeaker [2, 6]. Quantization noise is due to analog-to-digital conversion. These distortions cause a mismatch between the genuine and replayed speech pattern. This mismatch can be detected by training a classifier using cepstral-based or spectral-based fea- tures. According to the replayed speech threats, the ASVspoof2017 challenge is introduced concentrating on the replay spoofing attacks. This challenge aims to analyze vulnerability of ASV systems and its countermeasures in the face of replay attack [4]. In this study, we focus on the replay spoofing attacks and countermeasures based on the ASVspoof2017 dataset. In this work, we utilized six different features, Comprising Constant Q Cepstral Coefficient (CQCC), Modified Group Delay (MGD), Mel-Frequency Cepstral Coefficient …

Content truncated to 500 words