| Speaker recognition makes use of the speech coefficients which represent the speaker's voice feature to identify speaker, is a kind of biological certification technology. In recent years, speaker recognition widely draws the attention because of its convenience, efficiency and accuracy. It can be applied to a number of fields, is rapidly moving toward practical use, can widely be used in all walks of life.This paper is mainly about a text-independent speaker recognition system based on vector quantification (VQ) methods, a text-independent speaker recognition system based on Gaussian mixture model (GMM) and artificial neural network model. (ANN), we use LPCC and MFCC coefficient as the feature parameter set, respectively, in 15,25,41 of the speaker recognition library conducted experiments.Mainly works in this paper:(1) Feature extraction in some detail the feature extraction phase of the time-domain characteristics and LPCC, MFCC features, such as the extraction process, and different from the traditional DFT of the wavelet transform feature extraction.(2) It research into systems. The performance of various systems was studied. VQ model to study the code of the scale of the impact on system performance and the selection of threshold for the optimal design codebook some ideas put forward. In the GMM model, Gaussian mixture model to study the order of the number and length of training in voice impact on system performance, and the basis of a large number of experiments, the proposed voice-training of different length of the model given the recommended order; neural network in the trial of the experimental studies; an analysis of the advantages and disadvantages of various systems for hybrid recognition system created after the foundation.In the end, introduced the system and the various experiments, the results of the analysis and comparison of various parameters on the experimental settings are verified, the future prospect of research work carried out. |
PDF Full Text Request