Voice Recognition for User Authentication at Online Examinations

Abstract

Digitization has made a huge impact on education sector and many institutes all over the world had already transitioned to online from traditional face-to-face lectures. While learning is being practiced online, most of the examinations were conducted in general approaches within the institute’s premises. But the sudden outbreak of COVID-19 pandemic forced all the educators to adapt online platforms within a short period, who were being reluctant to shift earlier. Therefore, the necessity of digital assessment platforms with secure testing environments have arisen not only for educational sector but also in recruitment procedure of employees. In implementing an online examination system, the prime challenge is to maintain the credibility and the transparency with participants authentication. Therefore, introducing a better approach of user authentication is crucial in every examination platform. Considering the economical and the educational background in Sri Lanka, we are unable to expect the students to have accessibility of specific instruments and high-speed internet. Therefore, this study introduces a voice-based user authentication approach for online examinations which can be acquired with limited facilities. Voice based user authentication is to identify a user by analyzing the unique features of his /her voice sample. Recent studies show that user authentication with GMM (Gaussian Mixture Models) have efficiently used in speaker recognition. The key features of text independent voice signals are obtained using MFCC (Mel Frequency Cepstral Coefficients) and a unique model for the all the speakers who enrolled to the system is generated using Gaussian Mixture Models. The maximum likelihood algorithms are used to match the users voice samples against the speakers who are already enrolled. The dataset for the study is obtained from the UCSC/LTRL Speech corpus which contains 60 users with speech utterances of Sinhala language. A Web based application has been developed to implement the user authentication approach using python http (Hypertext transfer protocol) service and PHP. The accuracy of over 90% on correct identifications is obtained by models with voice samples relatively higher duration with GMM and MFCC where 20 trials are tested against the whole set of 60 trained models.