we got than

uh my name is

that was that in a little to and the to get that with the professor of S

from a you know that's

okay

and our team of

what what are you can

would like to present to a variable step-size the proposed and make a a type of in which we call

that

yeah or a solution for echo cancellation

uh

is the outline

as annotation

we can see

a what the production then uh

as the basis

of the proposed to eight of find projection

followed you followed by a

our proposed a and we and we

simulation up

company

now a sense to the nice presentation

yeah i

a scale but mission if we

if using production

basically we we deal with the same echo cancellation problem

we in fact is a

uh

system identification by with some challenge

this challenging in uh i have a union our maybe

is that a a long line

or the equal parts so a lot no think what they are not

if if you that and

maybe the double talk situations

see

is much

a challenging in there

a right

a this specific problem we

the the more give a task for some special Y

for example

if we want to

uh a faster

again and then choose the family of affine projection

oh

which we know that the

ah the

more efficient from this point of us compare

and that a man

and we also if we we won

to increase the robustness to double talk

we can use the body of a step size

we will see in the presentation

of course in general the as

a echo canceller

and keep with a little of detector but

yeah very useful

you is yeah that we've algorithm it somehow

oh was

but seven to that

and uh we have some of the hints let's say

uh

because course is the system we need to identify which are be

a a i i i

make is that the my is it

so it would be useful to the use

but to admit

i

adaptive of body

i i don't by means this three each with result

uh a the present a

now just the

oh

but if for the each of a you of the problem

uh this is the general

okay

separation

patient

what additional so

and the a lot of course

can be expressed in of a a ways for example from that use the point of view was the main

but

is to

but couple but C know the near a signal in the all yeah that

of course

for all a point of from the application point of view we have to cancel this K

well you not know what we have to like that if i is equal uh so this is the basic

and a the

application one

now i is a

uh talk uh and that the is the affine projection of what a a a a at may from the

convergence point of view

uh uh they but for very where especially we

the in it

signal inputs like speech

uh yeah creation

or the class

a find projection of are

you but yeah

uh

this is uh

it true signal vector but this as the input

signal signal my

and

most important this part i mean that you know the projection all that if you can see when P E

close to one

this is we used to a a lot and elements

this is a a a a if you and and

this is the step size for it that which we uh

the following

are those of its

H

now is that problem to me

a fine projection of what you can you thing by extending the idea of pnlms

in the same manner

as

we be form and lms the pnlms

we can do for

a yeah two

well it's it yeah

uh

it up yes this

that's a proper a marketing

which in fact is that i i i not automatic speech a lost to

a just just the step

for each the

in V but feel that way fish

if you can see here

let us denote by P of and is my

it would be very computationally efficient to compute a in the class that way

detection

we can take advantage of the i are gonna a kind of of this matter

and we can uh a process that it is yeah

this um

in this way

uh so here

B are not that sign

set

point

vector

is not the issue for a high number of do

so we can come

uh

so uh

P

and market

i simply multiplying

is a a a a a a a vector of the proper five

element by element with a vector or so

yeah these not means element by element multiplication between

these two but

as we can see

uh

the classical for more sound but they simple color the content

uh or what to make five

recently recently we proposed a soul call member

uh pop up to an the final projection of what

which

uh face into account not only the content

factors bar

that's that's study of these fact

and most important than

and uh

besides the fact that used are going to achieve

but that

performance in at instant

guess score that is that egg are i

it is more computation

efficient because

if you

take a look at C

uh

i

we need to compute or is the first

and N

the first

column

and to use

the out that form the previous

it that nation and and do the same

form for this map

this can be a a you know that if you should manner of my

computing or needs the first column and the first

which is not the case here

uh uh

you you the

it's uh

each of which you more plates

the uh

a

is that a on

uh medical many complex

that

patient

addition

uh there a class a lot of loading and our proposed a

of course

uh

yep and each becomes more smaller back and when we is the projection the

and uh

the next step

but what ones

which was to develop a of a step size

fashion for this

member in

uh uh how you

we we i here

uh uh it's a eight

and

again it is the step size mean

and is the we know

always

we have a complete think requirements

when we have to choose

the step size parameter because we have to compromise between combat

and misadjustment well

uh

the bit rate in

the double talk of who's that so

the you know that two

uh let's see

compromise is used the

and for out it yeah

it would be nice to use a body a step size the

and uh we start a lot development

by rewriting and it will be

this for a

uh using

uh a to use force the

step side

as as you can see

from these two

creations from this plot but it

if

we use

uh is the same but for all

and i and

we can get a each step size are working

which is

you can here

now

if we take a look

at a

a a posteriori a vector or not you are here a the which are you bring yeah

these these easy to uh what they no relation in these two that

using these two and a

a a of the seattle

now we can see is a relation and if we we remember that to the basic i

as you can find projection on reading once to chance and

be a posteriori here

the core we should

slides seas

assuming what was it

man

not well

we will get

a simple solution

which means that all the step size should be about one

unfortunately fortunately this holds already means absence of the nodes because is you remember from the

for

a

in the echo cancellation problem

we do not want to guess so

yeah

yeah are that if you don't by to recover is that near end signal

from these and so we modify

quantization

and we it i

uh

this is the

calculation that

that

then

we take the might or might have a

dictation

and the we get the

for

as a a a a a step size

of formal a like this

which yeah we have

the the elements of C L vector a

since is this is a very of but it is easy to estimate is this five

but unfortunately the problem but it mice

the estimate of the power

well of the near and

uh as the problems comes from the fact that the near end signal is in fact a combination P the

back noise and then you in speech

you when is two pass are uncorrelated heat so we can write as this relation

and even if we would basically to make this part and silence as for example

it is difficult to have an expression for this

uh

i

so

you know that to solve this problem

i i think that yeah that's that echo cancellation a configuration

so uh

we can express as you circulation in better

of

expectation

one and that of our

and now we will use a very strong assumption

which is that is yeah now if you that has come back somehow list and uh

get

in this case if if you make a a a a a a is this the assumption of this relation

can me

really then

axes

of course

as is this is a strong assumption because the

i'm i'm for example is the beginning of the adaptation one but i

station

it may not hold by

see E my the simulation that

so uh

oh performance a out of the fine

right would

uh uh most important

if we take a look at these the relation to this estimate all this

signal

all this uh measure are available because we only use

the signals

from the that if you that that that that is that i know and

it's all

oh from this point of view it would be a dark

a the solution

so the answer to our problem is

a state

step size of a space

step size

which look like to use of course

i i can be made

in the same or

say say that

and the finally

uh

we do show some uh

simulation results presented

you know a network echo cancellation problem

with a few that a lack of five and

the the wave actions

um the input will be you do white gaussian noise of speech

and a the the performance measure

a a a a a of the normalized means alignment or a according loss

a

and we will compare three at of buttons

our proposed or the last one

and

it's a uh uh a fixed stepsize action

uh uh as you can do these up here we do not mention anything about the alternate an eight five

you at can be we can was any what what's an eight back what

for

i

now uh you know uh i the experiment which is their i the elements of what in it's

what what and fact factor

was also presented by a monaural

before

uh in as

our simulation they'd

the bible

i mean

but is also

the the trans

and we compare

uh is this struggling in with

is that a have the step size

non proposed at that should the action

which is

uh

or they need to remove the

for puts an eight might from these

the ah

in the first

speed man

we compare the variable step that should with a fixed at signs

that's and was to

a fighting for the step size as we can see

uh as the body of the step size that's and combat almost as fast as

the

each step side that's start with a lot

size

but achieve a much lower miss i'm and close

and that

situation

in this case we use the

an input signal

uh

white gaussian noise

then

we uh are an snr all

uh

assuming that uh

try

patient the mean that

relation

and we compare our algorithm with a

none of course but may action

we can see that

i'll hope of that

is expected somehow

a of problems

a class L

action

in that of well convergence it

or that just

this alignment

and finally

uh uh you know that what looks and i not

uh

we compare at all

the three hour

and uh we see that for this point of view from that was to double a point of

our our that it's

much more what was as compare

yeah

at least is uh a non of a step size

yeah we use a a simple get about that that

a vector just two

and

the adaptation

double talk

i in it is a a a a few conclusion

so we propose

this out that point text of for echo cancellation

first

we can take advantage from a its computation of complex

T

which is low as from that is a class

second

we think that is important that the value of a step side

for

somehow somehow not about made

this means that we do not use any additional power uh

i'm sure one by one

and

we also note this is that

uh

these more what used to

uh near end signal evaluation

and most important to double

thank you very much earlier

for that

well we do you have some time for

i if anybody like

using again

you

presentation presentation just them curious about two

the step size it becomes negative when the values of the

design a value is a lot to than that of the there

yeah

so

in this problem

okay

this is a very nice for

because it we ski but it to meet some practical issues yeah

it is a recommended fast

to at

a small positive constant yeah

just to avoid you bytes

at most important

we should they is the absolute but

this a three star is then

the

but but what i

too much

but

if

inking your simulations used in order

a projection order

for for for yes

and is the method sense to the project no

no we just to choose this value

it's a for complex

but

we try was my

projection of

you have

and in your a personal experience

what projection order do you find is really the the right

and it to be working in

for for

uh

i think that uh to three we well we should not to take a larger than eight

that

but also because of the complexity issue you by that you for this these some uh

that's a diffusion map for completely because in a fine as the fine projection of what you

so main problem is how to invest

the out

i i have a project the higher projection of but we been blind or if you a quotation my

recently the as far as i

no was that are some more efficient manner

for example

you you all

because the models

i

or the might be end

variation of by

well also a

which she uh if you lead that for this operations so we can even go higher

a project

and other question

i think i have the opportunity to last one um

a more general question concerning the difference between acoustic echo cancellation

and network

and

could you clarify which

i think you first results at least as in the network

sure

yes

a the was the

so is the star

a point of

two and eight type of that if you that

oh

for a network that for consideration and respect to

that form but that's

send send you because of this i

of the

because i buy it can fall you by room acoustic a more cancellation if we properly choose

the proposed to eight five

this was that he's and defines the chose of the I the nlms out

which is

somehow of a host to this

ask

part think the

of course the

we can use for example a low as far as and uh as i

we proposed a few

a one year ago goes as the sparseness control coding which can

a beep on uh

well efficient as compared to ipnlms mess if we know as this from miss what it's to make some the

smell

so we is important to choose the proposed

well that's great thank you very much and