0:00:15so we first moved with allocation problem gmm-ubm systems
0:00:20the motivation is the biometric characteristics are not secret in the h increasing surveillance so
0:00:27think about you have a financial institution use all your one data there
0:00:31not the financial situation that was we line whose line and biometrics need to be
0:00:38projected on this protection it's be strong and from
0:00:41also to financial institutions got take that shouldn't be a lot that from one
0:00:45the data from one you cannot apply to the other issue transfer there
0:00:49just as an example and that's why we want to real research but what's biometric
0:00:55template protection for speaker recognition
0:00:58where you can go for biometric crypto system or can similar biometrics
0:01:04we wanted to
0:01:06emphasized on the properties of the ice was extended to four seven four five
0:01:11which is about biometric template protection and the basic things on there are very simple
0:01:15energy so i got the biometric one the data and i cannot which we've
0:01:20any thing right because it's to do much to not
0:01:24and unlike ability so i five and one system i cannot retrieve any information one
0:01:29of the system although they are based on the same ubm for example
0:01:33and what we propose a small bit allocation
0:01:38you can use what follows for the commitment or for divorce or schemes also for
0:01:43fast searches and basic idea is to extract common mfcc features and that supervector and
0:01:51then combined voice the supervectors
0:01:53and four button to template protection scheme using for example haven't distance comparison
0:02:01for improving the performance we also we did expected the most discriminant features so for
0:02:08that we use the relevant projection thank can see that post