| Speaker Recognition has increasingly become a hotspot of research in later years for it's a typical and important part of speech signal processing and has a wide range of applications which include banking or credit card transactions by phone, information and reservation services, access control in high security areas and forensic investigations. Speaker recognition deals with recognizing the identity of the person speaks, it is a research field of recognizing the speaker's identity on the basis of individual information included in the speech signals. It can be classified into speaker identification and speaker verification according to decision modes(one to many and one to one).With the development of communication and information technology, it is getting more and more attention for its bright future.The main research work and achievements are as following:First, I have done some work and researches in the field of front-end speaker recognition, and figured out a reasonable processing algorithm, then implemented it. Second, I have discussed some kinds of feature vectors, obtained a most valid feature vector: The Mel-frequency Cepstrum Coefficients (MFCC).Basing on the successfully extraction of MFCC, I have discussed the contribution of each coefficient to the final results. Third, in the training of speaker recognition models, I investigate the training of Gaussian mixture models (GMM).In this field, I utilized the Maximum Likelihood Estimate (ML) algorithm and Expectation-Maximization (EM) algorithm. Fourth, in the aspect of performance research, this paper has studied the performance of different numbers of Gaussian mixtures, in which the choice of mixture numbers related to training data are concluded. At the same time, several other parameters and factor were discussed and validated by experiments. Last, Multi-Threading technology was used to reduce the time of recognition. And a new method was used to advance the performance of the system when the speech library is very huge. |
PDF Full Text Request