Robust speaker verification system with anti-spoofing detection and DNN feature enhancement modules

Abstract

This thesis focuses on the robustness issues of speaker verification (SV) systems. Al-though current SV systems perform well under clean condition, their performance de-grades dramatically under real-world uncontrolled environments. The reliability of cur-rent SV systems is also questionable under spoofing attacks. These pitfalls severely limit it’s deployment in many applications. This thesis presents approaches to combat these two robustness issues, namely noise robustness and spoofing attacks. To address the noise robustness issue, the use of deep neural networks (DNN) as a feature compensation method in the front-end module of the SV system is proposed. The motivation to use DNN is due to its success in various related speech fields, and its ability to model nonlinear relationships between high dimensional input and output. In this work, DNN is used to convert noisy input features into clean features. The proposed method is evaluated using the benchmarking speaker recognition evaluation (SRE) 2010 dataset provided by the National Institute of Standards and Technology(NIST). To focus on the effect of feature pre-processing, the SV system is trained using noise free speech and evaluated on noise corrupted speech. Results show that the proposed DNN feature compensation improves the equal error rate (EER) by 2%-25% under different unseen noise types for various SNR levels. To address the spoofing attacks issue, the use of long temporal high dimensional speech features derived from both magnitude and phase spectra as input features to neural network (NN) classifiers is proposed. The long term temporal information is in-corporated by concatenating 31 successive frames as input feature to the NN classifier. The classifier is then used to predict the posterior probability of the test speech being spoofing speech. Four speakers of CMU-ARCTIC database are selected for spoofing data generation and methods evaluation. Spoofing data is generated by four synthesis meth-ods, namely: AHOcoder, STRAIGHT, JD-GMM with maximum likelihood parameter generation, and weighted correlation-based frequency warping (CFW). The results show that both long term information and detailed information maintained in high dimen-sional features improve the performance of synthetic speech detection significantly. The proposed method was extended and used to compete in the ASVspoof 2015 challenge and achieved best results in the closed set challenge among 16 teams worldwide.