Identification and authentication of user voice using DNN features and i-vector

Abstract

Currently, computerized systems, such as language learning, telephone advertising, criminal cases, computerized health care and education systems are rapidly spreading and creating an urgent need for improved productivity. Speech recordings are a rich source of personal, confidential data that can be used to support a wide variety of applications, from health profiling to biometric recognition. Therefore, it is important that the speech recordings are properly protected, so that they cannot be misused. The leakage of encrypted biometric information is irreversible and biometric links are renewable. The article proposes a block diagram of the identification of the users of the systems by individual voice characteristics, based on the joint use of the Deep Neural Network (DNN) method and $$i$$ -vector in the model of the elementary speech units, distinguished by increased security from various types of attacks on the biometric identification system, which allowed identifying the users with probability of first and second errors genus 0.025 and 0.005. The analysis of the vulnerability of the modules of the biometric voice identification system was performed and a structural scheme of the voice identification system of the user identification by voice with enhanced the protection against attacks was proposed. The use of elementary speech units in the developed identification systems makes it possible to improve computational indicators, reduce subjective decisions in biometric systems, and increase the security against attacks on the voice biometric identification systems.

Keywords

• voice identification
• voice authentication
• deep neural network
• speech recognition
• elementary speech unit (esu)