Siamese Networks for Audio Spoofing Attack Detection

Abstract

In recent times, audio spoofing attacks have become very common and serious in regard to Automatic Speech Verification(ASV) systems. There are many types of attacks possible like impersonation, speech synthesis, voice conversion or replay attack. Neural Networks mostly prove to perform well on such problems, and one of such is the Siamese Network which performs really well on simple data for classification. So in this paper we focused on detecting the audio spoofing attacks with the solution to it given by using Siamese Network. We have tried various different approach by combining different feature extraction techniques and various Siamese algorithms. We have used the popular Neural Network architecture, Siamese Networks which is hardly used in the audio domain. Audio data is difficult to work with, this method uses MFCC, Spectral Centroid, Spectrograms and Chroma Features for feature extraction, different python audio libraries like librosa, pyaudio, spafe etc. The result of this is combined with different Siamese networks and the performance is compared based on the Equal Error Rate(EER) of all these methods. So the method proposed in this paper is to use the very efficient Siamese Network algorithm for audio data and compare the performance of all variations and use it for detection of audio spoofing attacks.