The Research On Text-Independent Speaker Recognition Based On I-VECTOR

As a kind of biometrics,the speaker recognition is gradually accepted and used by the convenience of the use of convenience,non-interactive and so on,and become a research hotspot in the field of biometrics.The text-independent speaker recognition is to extract the information from the voice signal to reflect the characteristics of the individual,to complete the identification of the identity of the dialogue.In recent years,with the development of speaker recognition technology,speaker recognition has gradually moved towards social applications,but the actual use,due to the actual environment,the diversity of voice collection equipment and the length of the voice of the speaker and so on,so that the speaker identification There is still a problem in the recognition accuracy.In this dissertation,the i-vector model and the Gaussian linear discriminant analysis(GPLDA)model are studied from the perspective of compensation for the problems such as the low recognition accuracy and the environmental mismatch caused by the short voice in the actual use.The problem has been improved.The following is the work of this article.Firstly,this dissertation introduces the model of speaker recognition,discusses the pretreatment and feature extraction of speaker recognition,and uses the frequency cepstrum coefficient to extract the characteristics of the speaker.In view of the problem of insufficient training and test speech,GMM The model is analyzed,and the advantages and disadvantages of the system are analyzed.The advantages and disadvantages of the system are verified by experiment.The difference between the dynamic and static characteristics of the reaction speaker is studied.The influence of the characteristics on the speaker recognition,and the performance of the system is analyzed experimentally.Secondly,based on the characteristics of GMM-UBM cross-channel performance,based on the factor analysis,the i-vector based speaker recognition system is constructed by using the identity vector i-vector.For the problem of channel mismatch,Analysis and intraclass covariance normalization and other compensation means to compensate the system and analyze the impact of the compensation on the system.At the same time,the influence of i-vector dimension on speaker recognition system is analyzed by experiment,and the appropriate feature dimension is selected.Last,in this dissertation,Gaussian linear discriminant line analysis(GPLDA)model is used to transform i-vector into PLDA model for i-vector length normalization,resulting in the length of the normalized i-vector back-end covariance can not be accurately calculated,affecting the robustness of the system,this dissertation proposed the use of the whole variable space column vector length normalization Instead of normalizing the length of i-vector,and validating the proposed method,the results show that the method can improve the robustness of the system and the recognition rate is not reduced.