الفهرس 
Chapter 1:IntroductionOnly 14 pages are availabe for public view
 |  | from | 135 |  |  |
| | | from | 135 | | |
Abstract
Automatic Speaker Identification (ASI) is so crucial for security. It is utilized for many potential purposes such as voice Internet applications, telephone banking, security, surveillance, forensic science, and electronic system authentication. Current ASI systems perform well in generally quiet and clean surroundings. However, in noisy situations, the robustness of an ASI system against additive noise and interference is a crucial factor.
The interference can be assorted as stationary or non-stationary. Stationary interference is noise whereas non-stationary interference may be the voice from another speaker or music. Creating a reliable identification system in the presence of high levels of interference with a little dataset is the focus of this research. An investigation of the impact of interference on ASI system performance is presented in this thesis, which presents algorithms for achieving the most effective ASI system performance. The objective is to resist the interference of various forms.
This thesis presents two different trends for ASI system. The first trend is based on Artificial Neural Network (ANN) as a classifier and Mel Frequency Cepstral Coefficients (MFCCs) with pitch frequency as extracted features from speech signals. As part of the pre-processing step, various strategies are employed to boost system efficiency and minimize degradations. The second trend is based on deep learning with using a Convolutional Neural Network (CNN) that works on spectrograms and their Radon transformed images as inputs to increase the robustness of the ASI against interference effects. It is well-known that slight variations in images are not noticeable in their Radon Transform (RT). The simulation results indicate that the two proposed trends present satisfactory results. The model based on a CNN consumes less time than that needed for the ANN which requires many training epochs with classical hand-crafted features. Another issue that has been considered in this thesis is channel degradation elimination. Both noise and convolution effects have been considered even with interference. That is why both noise reduction and convolution techniques are adopted in this thesis as good solutions for the degradation effects.