Robust Speaker Recognition Using Speech Enhancement And Attention Model

Speech Application

Yanpei Shi, Qiang Huang, Thomas Hain

i or the u s and model

a target o k

the weighting rule

a problem in the unimodal

i all right

okay

i'm in the table

a little or it all i mean how men

and

to improve the performance you know i wrote you might not

it better

but the is better

we are currently are really hartman and recognition

okay tomorrow i in rate in one or more

i don't okay

you map or

a little more to the global mean not

a more or the u i

the problem we will have

i

we gotta identification

and the

a reasonable that the of course i

although

a new

notation o a whole number one

but i three right okay

general model unit

i write my

the other hand

or you might add more efficient model

probably not you know there

what the motivation for

no i nor my usual the problem all the speaker recognition

on the future

only we use when i

the are you speaker

nothing better than

a dimension

well i not information in

the core condition

really

well i

the mobile

okay well when i don't recognition model

the model

okay well

i

convolutional layer it more data

the speaker model

most people do you know well

we more

other than i'm recognition model

a the fact

independently

and to get

a little better than the one point five k u one and recognition

i mean that no the one that

it is not who and the in domain and with it

well

i

in all the speaker and you know

a whole and out in a

if you allusion and what it is you

well

of the movie

a more efficient addition there then you have the time

in addition there than l

well information open condition

you know the

the in the morning edition

a model adaptation okay really global condition

a eight conversation edition

and the

we have only okay

on the other hand recognition

i don't

a really and i will be a really by a all out of domain

well i and you the mac layer generating the n r

the calibration they the i

a binary okay or more

the little i

well by the channel estimate

on the other habitation

other application had available

a be a

but we directly are really apply channel

a where we have we are counting on timit

i between the core condition

the idea is to a you know not to buy or by two

i'm the weight

i really

using a not very i don't like to buy is

a visual word by

you the

the application i

i'm from the speaker who

we

at a

okay the handle because we have i

and

probably okay

the

enrollment data about

i o

i u

either the

a longer than that

where you that you are you

e o to o e

every meeting recognition

although the without a

and the

i know

i don't you do

and the other hand but

only one done by the okay

okay in that

well

well everyone to you to the

a e

and okay you did not

we have

the or

i

but i

the

speaker identification

you

the one approach

a that a whole training and that

i don't know there is no longer than the other hand but

a menu that and y two k

okay and that

okay

it

one okay

you know to do the whole problem

i don't know that will be used to a well known

did you are at the not and not you

are you that

only option

we you know

i but the i and i'll

i for the

but the it that

and you as a nice improvement

well

and i can be i

and maybe a data

a well you all the all pro

and there is a okay i'm weight and the

or you a bit

and i'm time and then the more

you have the right

and the other p

okay

and one speaker recognition

there are totally

really and then

okay

a fusion of it or not

well

and then than a the thing but okay not only

no value

a very

and that

i just a the university human

i'm not known

well

i

that

and it will be

the i e

we then

i'm getting higher than i would be other

and i three

a speaker verification and you the speaker recognition model only

i e

i e

are there

data a little

and the speaker recognition more

a using them

i right

i e

i four or more using a total and one iteration

well you

and i e

okay we have at

the i a model that using a total between and among the more you

the recognition

i'm the of the more like a you know how you know what

i am with organic probably layer

we don't well

i mean all

a little the

well

and a

okay the problem of the one level

i

it might have an error rate

a in well i don't really condition

and order to make the model i to make a movie that it would be

and the television if you

a it's a convolutional

well

okay

and then combine them here

there are more efficient

and the actual the model

and do not well what is that it may be a problem

well

i the them you

the we didn't use a suitable

but you wanna you know what

and that we the problem you

and the

but there is no

the

it'll

well

you the more you rate

i e

i performance of the extended edition

a

the united

and

i

i and i

and i e and i that the

and you want to than what we implement a one

a very good that you only not

how to i i e

and you know data model

she'll

by a nice improvement

you know a little you that'll

and i probability a really all you can

i

the information and okay

i

you know well

and then

but the

recognition

a lot of the

no

the

here

but not or you

from people i and so that

through a piranha

the individual model

if i the you know little or you can perform i might not

right

you can we try to model i s o a long

having to do not

them

relative improvement okay

in r e i e

i e at a

but the more

for the whole speaker recognition

a little i know the true value then problem the model

it completely okay on a

v you in a further more

there you go fishing model you years you model and a like i of it

you more normal or

then there is a little more i

the i

only the that

the biphone like

i don't you

okay i

you are

right

you tomorrow you are model

my anymore or and then

the women

information

and you have two more years

what do you all

and the three the original the real identification

and the

i i'm at

but i thirty and i

in a file

you know the a i e i'm sorry

i don't know that or a i u i e

of the plan

well

a mission it you know that only you the a i e and but i

e

and the barrel

the new combination the in narrative on the that

at the end and i

it is clear from the

that are not

a final no

bigger

i you know an additional

the video i e

if you're gonna i and

and i and i

a the i know

well you

i

you

i

the

recognition

although

on the number three

a the are really article

the level

and the by i i'm the colour

the p and

i didn't

on the paper

a lot of data by the

six hub four

and the unit for

well

you know the two

cool

the

identification

and that a simple

the or not

a possible

there is more

well you never the

all the model

and i

the relevant information

but the problem then you're able to spend the only you the graph model

and the over

okay

i four u o a i-vector model kind

what the

you great

or

and future work

that is that you the multi condition

i can be really you

and the what we're very well but

a movie that i

okay i four u

you