Automatic Speaker Verification

Abstract

We live in the digital era, protecting our confidential information is quite a difficult to deal with. Traditional method of securing the data like passwords or token is not enough. In addition to that more devices are connected to the Internet of Things, so the need of ironclad security is paramount. And this is where biometric system can help. Biometric authentication device uses behavioral characteristic of the person like fingerprint, voice, iris etc. So biometric authentication is an excellent way to secure the data. One such form of biometric is speaker verification and it is based on voice biometrics. Automatic speaker verification system checks that the speaker is the person who he or she claims to be. It is the process of verifying the identity on the bases of speech signal or voice print. There are two types of Speaker Verification System : Text-Independent Speaker Verification and Text-Dependent Speaker Verification. The former verify the speaker on the bases of speaker saying exactly enrolled or given words. And the other one verify the identity without the constraint of the speech. Text independent requires bigger training and testing utterances. Speaker verification falls into pattern matching problem. And many technologies used for processing and storing voice prints some of them are frequency estimation, hidden Markov models, Gaussian mixture models, pattern matching algorithms, neural networks, matrix representation, Vector Quantization and decision trees. Speaker verification depends on feature extraction and speaker modeling. There are many modeling techniques for speaker verification but here we have used joint factor analysis(JFA) modeling, UBM-GMM model and inter-session variability(ISV) modeling, i-vector cosine modeling, i-vector-PLDA modeling and artificial neural network with two different data-sets. Here experiments are done with bob spear toolkit with different speaker modeling techniques on voxforge dataset. Automatic speaker verification is vulnerable to different spoofing attacks so here, different spoofing attacks and their countermeasures are discussed. The work is done on replay attack with different feature fusion techniques on ASVspoof2017 dataset.