Font Size: a A A

Design And Implementation Of A Speaker Recognition System

Posted on:2015-01-14Degree:MasterType:Thesis
Country:ChinaCandidate:B LiuFull Text:PDF
GTID:2308330479479201Subject:Electronics and Communications Engineering
Abstract/Summary:PDF Full Text Request
Nowadays, information security has become increasingly concerned as the highly developed information technology. Convenience, economy and accuracy of biological recognition make it to be the most focused part in research and development of information security.Speaker recognition, an important part of biological recognition technology, estimates and identifies the user via extracting and analyzing matching of speech features. For decades, speaker recognition has been further developed and progressed since the persistent efforts of researchers.This paper focuses on the study and analysis of the design and realization of speaker recognition. The author designs a speaker recognition system, which tests and analyses the performance of speaker recognition system in different dimensions, number of speakers and time, feature parameters and endpoint through the principle analysis and simulation realization of each module of the system. The concrete work of this paper is organized as follows.First, analyzing and researching the principle and application of window function and voice signal endpoint detection. The author compares performance of different functions and methods to demonstrate its feasibility verified by experiments, and propose a new method based on chaotic character for endpoint detection. The experiments show that the method can be better at excluding noise to distinguish noise and voice signal.Second, analyzing and researching the principle and realization of feature extraction of voice signal and recognition of speakers. The author analyzes and compares the performance of several characteristic parameters impacting the system validated simulation experiments. The author studies the principle and realization of LPCC, MFCC, and then proposes a novel feature parameter of MFCC and differential MFCC, and this dynamic feature parameter of MFCC get better performance than the others. The author studies fixed-point realization of floating-point in the process of extraction of MFCC feature coefficient. The author tests its accuracy and precision and researches how it affects performance of the system. Then the author studies and analyzes the GMM(Gaussian Mixture Model) and vector quantization which are the basic principles and processes of speaker recognition. Two experiments are conducted to research the performance of recognition algorithm.Finally, designing a GMM-based speaker recognition system at the platform of Matlab. The system is tested in a variety of conditions of data and analyzed in theory, and the author finds that it almost meets the demand of performance of engineering.
Keywords/Search Tags:speaker recognition, Gaussian mixture model, Embedding dimension, Mel-frequency cepstral coefficent
PDF Full Text Request
Related items