Model of the speaker identification and verification subsystem

Анотація

 The paper is focused on the pressing problem of authentication and verification of speakers based on voice information, which plays an important role, for example, in online or remote communication and information exchange in all spheres of life, including scientific communication. The aim of this paper is to create a model of a speaker identification and verification subsystem. To achieve this goal, the following tasks were accomplished: the connection of the modules of the proposed model was explained, the voice information analysis module was explored, while ensuring the scalability of the system with a significant increase in the number of users, and the results were analyzed. The developed pseudo-ensemble-based neural network module was tested on a dataset prepared on the basis of the LibriSpeach corpus, an open English speech corpus based on the LirbiVox project of voluntarily provided audio books. The result of applying the developed module on the selected dataset is demonstrated, demonstrating that in order to implement the subsystem in a neural network training system, the proposed pseudo-ensemble should be trained on at least 120 epochs using noise reduction methods at the stage of audio sequence preprocessing.