Voice feature extraction and speaker recognition using discrete signal processing & neural networks

Dissertant

Daghbosheh, Muhammad Ayyad

Thesis advisor

Hattab, Izz al-Din Shakir Hasan

Comitee Members

al-Shaykh, Asim A. R.
al-Sarayirah, Bashshar
al-Lahham, Muhammad Ismail Abd al-Rasul

University

Arab Academy for Financial and Banking Sciences

Faculty

The Faculty of Information Systems and Technology

Department

Computer information systems

University Country

Jordan

Degree

Ph.D.

Degree Date

2012

English Abstract

The use of biometric information has been known widely for both person identification and security applications.

It is common knowledge that each person can be identified by the unique characteristics of one or more of person biometrics.

One of the biometric characteristics of that a person can be identified by his voice.

This research attempts to find the best discrete coefficient that can be selected from one or more voice level using neural network as classifier.

The proposed speaker recognition system consists of three phases; preprocessing phase (two processes are performed on the sound, DC level removal and resize of sample for 2000 sample for each sound), feature extraction phase (features that distinguish each sound from another, in this research discrete wavelet transform is used for extracting features from sounds), and recognition phase (many classifiers could be used for speaker recognition, in this research supervised neural networks, MLP and LVQ are used as classifiers.

This research is concerned with studying the feature extracted from discrete wavelet transformation using feed-forward MLP and learning vector quantization as classifiers, the neural networks trained with features extracted from different levels of discrete wavelet transform (one level at a time).

The research illustrates the effected of using different percent (50 %, 40 %, 30 %, 20 %, 10 %) of each level instead of all feature, then compare their recognition ability and decide the best level in addition to the percent of features that are enough to give comparable results with respect to all feature set and decide the neural network architecture that gives the best recognition rate.

The data set consists of different sounds recorded from ten different persons (three female and seven males).

The system was trained and tested using cross-validation.

Main Subjects

Information Technology and Computer Science

Topics

• Digital communications
• Digital electronics

No. of Pages

104

Table of Contents

• APA
• MLA
• AMA

Language

English

Data Type

Arab Theses

Record ID

BIM-306668