The Research On Technology Of Building A Speaker Recognition Database

Speaker recognition technology has been widely used in our daily life.With the development of mobile network,speech data is increasing rapidly.And so,research of fast recognition in large scale corpus has been paid high attention now.However,not only the recognition algorithm itself is being a challenge,but how to build a sizeable speaker recognition database really matters.Traditional building strategies have high cost and long period,it can hardly meet the need above.We managed to use the Internet as a retrieve target and then designed a speaker recognition database building method based on speaker segmentation and clustering technology.This method can lower the cost in speech retrieve step,and can also reduce the time cycle by a semiautomatic execution.We studied core parts like speech execution and speech verification in the building process.In speech execution part,we made improvements in voice activity detection,distance in clustering and speaker modeling.We proposed a new sub-band entropy detection feature based on prior reasearch,and the experiment showed that it can obtain great robustness in variable noisy environments.Combined data feature with applied purpose,we designed a two stage clustering algorithm,which use T2 distance as the first criterion and information difference as the second judgement.This algorithm resulted in a better cluster purity eventually.As speaker modeling,after observing and analyzing error instances,we use the difference between background likelihood score and local model likelihood as a criterion.And then delete invalid or bad speech frames.In speech verification part,in order to delete overlapped speech files and filter good files,we designed a speech check and speech filter mechanism based on speaker verification technology.Combined with the above methods,we designed a speaker recognition databased building method.Through experiment results,we can see the improvement of proposed algorithms.At last,we used this method built a speaker recognition database that contains 18833 speakers.And then analyzed its usability by sampling survey.