Impostor Recognition Based Voice Authentication by Applying Three Machine Learning Algorithms

Abstract

As compared to other conventional biometrics systems, voice is a unique and important metric, where it is used in many vital fields as the security and communication domains that do not need to be expensive to achieve. The purpose of this article is to see how machine learning (ML) algorithms perform for speaker Authentication to recognize impostors. To boost the audios usable in real environments, it was suggested the preprocessing of audio, like noise decreasing and voiced improving. Mel Frequency Cepstral Coefficients (MFCC) and the four features (Amplitude, Zero Crossing, Mean, and Standard Division) are extracted for all audio metrics, straight beside their differentials and accelerations. Then, Vector Quantization (VQ) is applied to these files. The algorithms were prepared and examined on two datasets, by applying k-fold cross-validation. The preparation for testing and comparing the three (ML) approaches are as follows: Support Vector Machine (SVM), One Rule (One-R), Linear Regression (LR). The result of the (SVM) algorithm average accuracy of 96.33 percent was superior. Index Terms— ML, Speaker Authentication, SVM, LR, ONE.R, Impostor.