without someone so the world that i'm going to present so you today is and

extend personal for one of the sub subsystems so that the ice quality man

submitted form needs the elderly and the false and fifteen challenge

and the

although it is focused on delivery a fifteen we believe that the key components or

four

in this of this paper they can be used in a much wider context

so the first thing that we explore is that the how to more get the

most of the key lda for

and discriminant for discriminative training

because nist ovaries that close the set identification task and it's going to train it

should be the best

so the most important thing is that we show that if we if we're

and i u

if we used plp parameters to projects i-vectors on the p lda latent subspace apply

it discriminative methods in that subspace and then project them but

then a week and improve the performance compared to just the baseline the one

we use lda and then

and maximum mutual information of their from on top of it

and the second the

important thing is that the we show how four

take

lre

cost function

approximated and use it as an objective function for a discriminative l the difference

so that as a can see

it is based on false acceptance rate and false rejection rate and all of them

use indicator functions so we approximate those functions

into continues functions and then

where able to differentiate them

and the of course of these method in general you can be used for any

in a cost function in theory

okay thank you for attention to the