The Designing And Implementing Of A Neural Network Based Speaker Recognition System

Voiceprint Recognition, also called Speaker Recognition (SR), is a new bio-recognition technology raised in recent years. Voiceprint Recognition has attracted people's attention by its convenience, economically and accuracy. By the development of computer technologies in recent years, Voiceprint Recognition also has gotten enough development. It has been applied in many areas such as police detection, voice controlling, diagnosing and so on. Voiceprint Recognition identifies speakers by speech features from speech waves. Different from Speech Recognition, Voiceprint Recognition does not concern with the contents of speech signal, but the speaker's features. Voiceprint Recognition has two key technologies: first one is feature extraction, and the second one is the recognition model.This paper describes the principle and general procedure of Voiceprint Recognition and several problems:1. The extraction of features of Voiceprint Recognition System. The paper describes the sound channel model, Linear Predictive Coding (LPC), Linear Predictive Cepstral Coefficients and Mel-Frequency Cepstral Coefficients.2. The recognition models of Voiceprint Recognition System. Hidden Markov Model, Gaussion Mixture Model, Vector Quality Model and Support Vector Machine had been introduced in this paper.Based on the studies of basic principles and technologies, we have designed a Voiceprint Recognition prototype system using Artificial Neural Network (ANN) as the recognition model. The system uses LPCC and MFCC as the basic features. The computation complexity of MFCC parameter is low, and LPCC parameter is also effective in SR, so we selected MFCC and LPCC as the main feature in our paper. The optimization of Neural Network parameters is the key problem of NN, most researches use GA or BP algorithms to optimize the NN parameters. This paper uses Particle Swarm Optimization (PSO) algorithm to optimize the NN parameters. PSO is a group intelligence technique. Compared with GA, PSO has an easier coding method, and the algorithm is easier to realize and comprehend.By using the Voiceprint Recognition prototype system, we did plenty of experiments in laboratory circumstance. These experiments gathered source voices by all kinds of methods, such as context irrelevant and context relevant voices. We also compared and analyzed these experiments. Experiment results have shown that PSO and ANN are effective in Voiceprint Recognition.