Design And Implementation Of Phrase Dependent Speaker Recognition System

Speaker recognition is a research focus in the field of speech recognition,and it is widely applied in many scenes.After a long-term research,the techniques applied in speaker recognition have been developed,meanwhile the application of Joint Factor Analysis(JFA)in speaker modeling has achieved great success.In practical application,in order to make the recognition more convenient,each utterance of each speaker can not be too long.In order to prevent the speakers' passwords from being stolen,the content of the phrase can not be fixed.Thus,we proposed a phrase dependent speaker recognition system.Compared with the traditional text independent speaker recognition system,text dependent speaker recognition system has a wider range of applications.This paper is carried out from the following aspects:(1)On the basis of familiarity with the overall architecture of speaker recognition,we make a deep study of the theory and key technologies of Joint Factor Analysis.Besides,we design and implement the speaker recognition system based on JFA and compare several schemes in each part of the systems;(2)We design and implement a phrase dependent speaker recognition system,especially make a deep study of the speaker verification.For the phrase dependent part,we propose a method which can select the most diverse text combination for each speaker on the use of Silhouettes.This method improves the performance of the system.Aiming at the part of threshold setting,the usual threshold setting methods are summarized.And this paper also proposes a new confirm threshold setting method using the Silhouettes,which is efficiently applied in the dual threshold decision mechanism;(3)This paper proposes a group based decision-making mechanism for speaker verification and makes a comparison between it and the traditional decision-making mechanism.This mechanism can find a similar speaker group for each register which can assist the finally decision making.