Deep Learning Algorithms Based Voiceprint Recognition System in Noisy Environment

Abstract

Voiceprint Recognition (VPR) is the mechanism by which a user's so-called identity is determined using characteristics taken from their voice, where this technique is one of the world's most useful and common biometric recognition techniques particularly the fields- relevant to security. These can be used for authentication, monitoring, forensic identification of speakers, and a variety of related activities. The aim of this thesis is to design a deep learning strategy, which will provide a way to implicitly learn the voiceprint recognition in noisy environments.Two approaches were used for VPRS and were used the same structure of convolution neural network for two approaches and trying to increase the accuracy of the system by deal with a huge dataset adding to its background a random noise to prove the efficiency of the system in noisy conditions. In this thesis, Attempt is applied to create a system that recognizes human speaker identity using Convolutional Neural Network (CNN). Used CNN for both feature extraction and deep learning algorithm,thus will enhance the ability of the system to be much accurate and be more efficient. The CNN architecture is designed to work with both MFCC-CNN and RW-CNN. In both cases, the proposed CNN inputs are images, i.e. the network dealt with images, where the same CNN architecture is used for both methods .The obtained findings show that both methods are similar in their accuracy 0.96 and mean square error 3.2000e-08 results but differents in performance where the time results show that RW-CNN is better than MFCC-CNN whether with or without noise. In other words RW-CNN is more efficient in clean and noisy environments from MFCC-CNN.