User Verification Based On Customized Sentence Reading

Speaker verification systems have been one of the most important means for user authentication on modern devices and systems such as smartphones, smart buildings, etc. Existing speaker verification systems are prone to replay attacks due to the acoustic nature of human voices. In this paper, we aim at designing a novel speaker verification system which is immune to replay attacks while achieving accurate verification results. We propose a sentence-level customization scheme, which generates different types of sentences for speaker training. The training and testing voice data is different but within the same category, which naturally avoids replay attacks and reserves the ability to identify speakers. Experiments will be conducted on both RSR2015 and NIST SRE10 databases in our future work.