The Research And Implementation Of Speaker Recognition System Based On Vector Quantization

Speech is not only the most frequently used and the most convenient tool of communication, but also the main way for us to get access to information. The speaker recognition which is an important branch in the field of speech signal processing, is a technique to identify the speaker by speech signal. This technique is mainly used in the field of security, judicature and so on.This paper builds a set of speech recognition software system based on a template training method of vector quantization. The main contents are as follows:(1) Different methods of the speech signal pre-processing and feature extraction are elaborated in this paper. The effects of the rectangular window and hamming window on speech signal analysis are studied. Three methods of endpoint detection including double threshold method, spectral subtraction method and spectral entropy method are compared. Extraction methods of LPC, LSP, LPCC, MFCC, ALPCC and AMFCC are introduced. The basic principles, process and algorithms of vector quantization are elaborated.(2) With Matlab GUI, M language and compiler for tools, the speech recognition software system is set up. In front of the system, the login module and speech database management module are set up to provide a safe and convenient using environment. In the main interface of the system, there are mode selection module and speaker recognition process modules that contain all the ways introduced before. Features are:the system has many kinds of alternative programs, and users can set the parameter values by themselves. Besides, both the identification of a single person and the identification of more than one person can be carried out. The system has visual effects, and its operation is very simple. There are feedback information module and other functional modules such as speech time domain analysis module, speech transform domain analysis module in the main interface. They expand the use range of the system and make it more flexible.(3) System testing and experimental analysis. Testing show: the modules are well-coordinated, and can complete the task of identification successfully. Three experiments are carried out to study the influence of characteristic parameters and order, the size of codebook and the size of frame and shift on recognition rate. Results show:AMFCC with14order has the best recognition rate, under the conditions of the size of codebook is128and the size of frame and shift are440and220.