Voice Based Authentication Using Mel-Frequency Cepstral Coefficients and Gaussian Mixture Model

Abstract

Voice operated devices are becoming popular nowadays. For this it is necessary that voice authentication is secure. In this paper, we address some known attacks like replay, personification and attacks using AI voice bots and limitations like text and language dependency of human voice authentication systems. We have also developed an interactive system to tackle these problems. The system verifies the user by performing voice matching as well as on an intellectual level by asking questions which only humans are able to answer and not any AI bot. In the system, an average user requires around 35 seconds for registration and around 25 seconds for authentication. The system's accuracy comes out to be 97.8% for English speakers and 95% for Hindi speakers.