Design and Implementation of a Stand-alone Voice Recognition System

Abstract

For the past several decades, the biometric security systems varied between those based on behavioral and others based on physiological features of people. One of the behavioral biometric security systems is that based on voice features of people and is called speaker recognition system. As a result of the advance of machine learning and computer technology, speaker recognition has rapidly evolved and has become very popular in the recent years. It is being intensively researched and found many applications where it saves lot of troubles which appear as a result of using other biometric security systems, so it is considered a high security system. Many methods were presented for the aim of designing an automatic speaker recognition system and other methods were presented for the aim of developing reliable speaker recognition systems. In general, any speaker recognition system involves four basic steps: a) data base formation, b) pre-processing, c) features extraction, and d) classification or matching of extracted features. In this work we have developed a method for speaker recognition using the English Language Speech Database for Speaker Recognition (ELSDSR) database which is compose of audio files for training and others for testing. The developed method starts by pre-processing the training audio files of the database. Then the Wavelet Packet Transform (WPT) is employed on the pre-processed files for feature extraction purposes. For the excessive number of features provided with the WPT, the energy corresponding to each WPT node is calculated to reduce the dimensionality of the wavelet coefficients by removing redundant features and to form features vectors. The features vectors are sent to the Feed Forward Back-propagation Neural Network (FFBPNN) system.