Audio-replay Attacks Spoofing Detection for Automatic Speaker Verification System

Abstract

Speaker verification, as one of the most convenient methods in biometric systems, has been widely applied. However, most automatic speaker verification systems are vulnerable to a variety of attacks, especially audio-replay attacks. Therefore, this paper focuses on the detection method of audio-replay attacks. It is found that using a hybrid feature can achieve better results than a single feature. As for the classifier in the backend, the use of the Gaussian mixture model cannot obtain convincing results, and the traditional convolution neural network architecture can easily lead to the model overfitting. Therefore, this paper chooses DenseNet architecture. The experimental results show that the hybrid feature and DenseNet architecture can achieve 46.06% relative improvement than the baseline system.