Speaker recognition for security systems under noise effects

Nowadays, many different devices and applications such as vehicles, smart home devices, mobile banking, automatic dictation programs, and legal surveillance comprise speech and Speaker Recognition (SR) systems. Noise is one of the most important factors that affect the performances of these systems. Therefore, reducing the susceptibilities of the systems to noise is very important. Two different methods are proposed within this thesis to reduce the negative effects of the noises for SR systems. One of these methods is creating impostor models by clustering speaker models. The other method is a Polynomial Regression (PR) based Voice Activity Detector (VAD), which aims to determine the high energy speech regions under additive noise. Recent, and widely used SR methods, and the proposed algorithms within this thesis were realized experimentally, and performance analyzes were made by comparatively presenting results of the case studies.