Speaker Verification System for Online Education Platforms

摘要

Online education is becoming more popular every day as it provides many advantages over traditional education. However, it also comes with several challenges. In terms of security, the challenge lies in verifying the identity of the student, to prevent frauds. This work analyzes the use of i-vectors to perform speaker verification in an online education scenario. A database with speech signals from 103 speakers and speech tasks in two different languages was collected for the present experiment. The proposed method comprises the following steps: (1) acoustic features of the signals are extracted, (2) an i-vector extractor is trained and used to, (3) extract i-vectors from test signals, (4) the extracted i-vectors are processed with PCA whitening and PLDA scores obtained, and (5) the evaluation metrics are computed from these scores. The results show the best Equal Error Rate (EER) at 10.3% using speech signals in the native language of the speakers.