0:00:13we got than
0:00:14uh my name is
0:00:15that was that in a little to and the to get that with the professor of S
0:00:20from a you know that's
0:00:22and our team of
0:00:24what what are you can
0:00:26would like to present to a variable step-size the proposed and make a a type of in which we call
0:00:33yeah or a solution for echo cancellation
0:00:38is the outline
0:00:39as annotation
0:00:41we can see
0:00:42a what the production then uh
0:00:45as the basis
0:00:47of the proposed to eight of find projection
0:00:49followed you followed by a
0:00:51our proposed a and we and we
0:00:53simulation up
0:00:56now a sense to the nice presentation
0:01:00yeah i
0:01:02a scale but mission if we
0:01:04if using production
0:01:06basically we we deal with the same echo cancellation problem
0:01:09we in fact is a
0:01:12system identification by with some challenge
0:01:15this challenging in uh i have a union our maybe
0:01:19is that a a long line
0:01:21or the equal parts so a lot no think what they are not
0:01:24if if you that and
0:01:25maybe the double talk situations
0:01:29is much
0:01:29a challenging in there
0:01:31a right
0:01:33a this specific problem we
0:01:35the the more give a task for some special Y
0:01:38for example
0:01:39if we want to
0:01:41uh a faster
0:01:43again and then choose the family of affine projection
0:01:47which we know that the
0:01:49ah the
0:01:50more efficient from this point of us compare
0:01:53and that a man
0:01:54and we also if we we won
0:01:56to increase the robustness to double talk
0:01:59we can use the body of a step size
0:02:01we will see in the presentation
0:02:03of course in general the as
0:02:05a echo canceller
0:02:06and keep with a little of detector but
0:02:09yeah very useful
0:02:10you is yeah that we've algorithm it somehow
0:02:13oh was
0:02:13but seven to that
0:02:16and uh we have some of the hints let's say
0:02:21because course is the system we need to identify which are be
0:02:25a a i i i
0:02:27make is that the my is it
0:02:29so it would be useful to the use
0:02:32but to admit
0:02:34adaptive of body
0:02:36i i don't by means this three each with result
0:02:39uh a the present a
0:02:45now just the
0:02:48but if for the each of a you of the problem
0:02:50uh this is the general
0:02:54what additional so
0:02:56and the a lot of course
0:02:57can be expressed in of a a ways for example from that use the point of view was the main
0:03:03is to
0:03:04but couple but C know the near a signal in the all yeah that
0:03:09of course
0:03:10for all a point of from the application point of view we have to cancel this K
0:03:15well you not know what we have to like that if i is equal uh so this is the basic
0:03:20and a the
0:03:22application one
0:03:25now i is a
0:03:27uh talk uh and that the is the affine projection of what a a a a at may from the
0:03:33convergence point of view
0:03:35uh uh they but for very where especially we
0:03:38the in it
0:03:39signal inputs like speech
0:03:41uh yeah creation
0:03:43or the class
0:03:44a find projection of are
0:03:46you but yeah
0:03:51this is uh
0:03:52it true signal vector but this as the input
0:03:54signal signal my
0:03:56most important this part i mean that you know the projection all that if you can see when P E
0:04:02close to one
0:04:03this is we used to a a lot and elements
0:04:06this is a a a a if you and and
0:04:09this is the step size for it that which we uh
0:04:12the following
0:04:14are those of its
0:04:16now is that problem to me
0:04:18a fine projection of what you can you thing by extending the idea of pnlms
0:04:24in the same manner
0:04:26we be form and lms the pnlms
0:04:29we can do for
0:04:30a yeah two
0:04:31well it's it yeah
0:04:35it up yes this
0:04:36that's a proper a marketing
0:04:39which in fact is that i i i not automatic speech a lost to
0:04:43a just just the step
0:04:45for each the
0:04:46in V but feel that way fish
0:04:49if you can see here
0:04:51let us denote by P of and is my
0:04:53it would be very computationally efficient to compute a in the class that way
0:04:59we can take advantage of the i are gonna a kind of of this matter
0:05:03and we can uh a process that it is yeah
0:05:07this um
0:05:09in this way
0:05:11uh so here
0:05:13B are not that sign
0:05:18is not the issue for a high number of do
0:05:20so we can come
0:05:23so uh
0:05:25and market
0:05:26i simply multiplying
0:05:28is a a a a a a a vector of the proper five
0:05:31element by element with a vector or so
0:05:35yeah these not means element by element multiplication between
0:05:39these two but
0:05:41as we can see
0:05:43the classical for more sound but they simple color the content
0:05:47uh or what to make five
0:05:51recently recently we proposed a soul call member
0:05:55uh pop up to an the final projection of what
0:05:58uh face into account not only the content
0:06:02factors bar
0:06:03that's that's study of these fact
0:06:05and most important than
0:06:07and uh
0:06:09besides the fact that used are going to achieve
0:06:12but that
0:06:12performance in at instant
0:06:14guess score that is that egg are i
0:06:17it is more computation
0:06:19efficient because
0:06:20if you
0:06:21take a look at C
0:06:24we need to compute or is the first
0:06:27and N
0:06:28the first
0:06:30and to use
0:06:31the out that form the previous
0:06:32it that nation and and do the same
0:06:34form for this map
0:06:36this can be a a you know that if you should manner of my
0:06:40computing or needs the first column and the first
0:06:42which is not the case here
0:06:47uh uh
0:06:48you you the
0:06:49it's uh
0:06:49each of which you more plates
0:06:51the uh
0:06:54is that a on
0:06:55uh medical many complex
0:06:59uh there a class a lot of loading and our proposed a
0:07:03of course
0:07:06yep and each becomes more smaller back and when we is the projection the
0:07:13and uh
0:07:14the next step
0:07:15but what ones
0:07:16which was to develop a of a step size
0:07:19fashion for this
0:07:21member in
0:07:22uh uh how you
0:07:24we we i here
0:07:25uh uh it's a eight
0:07:29again it is the step size mean
0:07:31and is the we know
0:07:35we have a complete think requirements
0:07:38when we have to choose
0:07:39the step size parameter because we have to compromise between combat
0:07:44and misadjustment well
0:07:46the bit rate in
0:07:48the double talk of who's that so
0:07:50the you know that two
0:07:52uh let's see
0:07:53compromise is used the
0:07:56and for out it yeah
0:07:57it would be nice to use a body a step size the
0:08:02and uh we start a lot development
0:08:05by rewriting and it will be
0:08:07this for a
0:08:09uh using
0:08:11uh a to use force the
0:08:13step side
0:08:15as as you can see
0:08:16from these two
0:08:18creations from this plot but it
0:08:21we use
0:08:22uh is the same but for all
0:08:25and i and
0:08:26we can get a each step size are working
0:08:29which is
0:08:30you can here
0:08:33if we take a look
0:08:34at a
0:08:36a a posteriori a vector or not you are here a the which are you bring yeah
0:08:41these these easy to uh what they no relation in these two that
0:08:46using these two and a
0:08:49a a of the seattle
0:08:51now we can see is a relation and if we we remember that to the basic i
0:08:56as you can find projection on reading once to chance and
0:09:00be a posteriori here
0:09:01the core we should
0:09:02slides seas
0:09:04assuming what was it
0:09:06not well
0:09:07we will get
0:09:08a simple solution
0:09:10which means that all the step size should be about one
0:09:14unfortunately fortunately this holds already means absence of the nodes because is you remember from the
0:09:21in the echo cancellation problem
0:09:23we do not want to guess so
0:09:25yeah are that if you don't by to recover is that near end signal
0:09:29from these and so we modify
0:09:32and we it i
0:09:35this is the
0:09:36calculation that
0:09:39we take the might or might have a
0:09:43and the we get the
0:09:45as a a a a a step size
0:09:48of formal a like this
0:09:50which yeah we have
0:09:52the the elements of C L vector a
0:09:55since is this is a very of but it is easy to estimate is this five
0:09:59but unfortunately the problem but it mice
0:10:02the estimate of the power
0:10:04well of the near and
0:10:06uh as the problems comes from the fact that the near end signal is in fact a combination P the
0:10:11back noise and then you in speech
0:10:14you when is two pass are uncorrelated heat so we can write as this relation
0:10:19and even if we would basically to make this part and silence as for example
0:10:24it is difficult to have an expression for this
0:10:30you know that to solve this problem
0:10:33i i think that yeah that's that echo cancellation a configuration
0:10:38so uh
0:10:41we can express as you circulation in better
0:10:46one and that of our
0:10:48and now we will use a very strong assumption
0:10:52which is that is yeah now if you that has come back somehow list and uh
0:10:58in this case if if you make a a a a a a is this the assumption of this relation
0:11:03can me
0:11:03really then
0:11:06of course
0:11:07as is this is a strong assumption because the
0:11:10i'm i'm for example is the beginning of the adaptation one but i
0:11:16it may not hold by
0:11:18see E my the simulation that
0:11:20so uh
0:11:21oh performance a out of the fine
0:11:24right would
0:11:25uh uh most important
0:11:27if we take a look at these the relation to this estimate all this
0:11:32all this uh measure are available because we only use
0:11:35the signals
0:11:36from the that if you that that that that is that i know and
0:11:40it's all
0:11:41oh from this point of view it would be a dark
0:11:44a the solution
0:11:45so the answer to our problem is
0:11:48a state
0:11:49step size of a space
0:11:50step size
0:11:51which look like to use of course
0:11:53i i can be made
0:11:55in the same or
0:11:56say say that
0:11:58and the finally
0:12:02we do show some uh
0:12:04simulation results presented
0:12:06you know a network echo cancellation problem
0:12:09with a few that a lack of five and
0:12:11the the wave actions
0:12:13um the input will be you do white gaussian noise of speech
0:12:18and a the the performance measure
0:12:21a a a a a of the normalized means alignment or a according loss
0:12:26and we will compare three at of buttons
0:12:31our proposed or the last one
0:12:34it's a uh uh a fixed stepsize action
0:12:38uh uh as you can do these up here we do not mention anything about the alternate an eight five
0:12:43you at can be we can was any what what's an eight back what
0:12:49now uh you know uh i the experiment which is their i the elements of what in it's
0:12:56what what and fact factor
0:12:57was also presented by a monaural
0:13:01uh in as
0:13:02our simulation they'd
0:13:04the bible
0:13:04i mean
0:13:05but is also
0:13:06the the trans
0:13:08and we compare
0:13:09uh is this struggling in with
0:13:11is that a have the step size
0:13:13non proposed at that should the action
0:13:16which is
0:13:18or they need to remove the
0:13:20for puts an eight might from these
0:13:22the ah
0:13:24in the first
0:13:24speed man
0:13:25we compare the variable step that should with a fixed at signs
0:13:29that's and was to
0:13:31a fighting for the step size as we can see
0:13:34uh as the body of the step size that's and combat almost as fast as
0:13:40each step side that's start with a lot
0:13:43but achieve a much lower miss i'm and close
0:13:46and that
0:13:49in this case we use the
0:13:51an input signal
0:13:54white gaussian noise
0:13:56we uh are an snr all
0:14:01assuming that uh
0:14:04patient the mean that
0:14:06and we compare our algorithm with a
0:14:08none of course but may action
0:14:11we can see that
0:14:12i'll hope of that
0:14:13is expected somehow
0:14:15a of problems
0:14:16a class L
0:14:18in that of well convergence it
0:14:22or that just
0:14:23this alignment
0:14:24and finally
0:14:26uh uh you know that what looks and i not
0:14:30we compare at all
0:14:32the three hour
0:14:34and uh we see that for this point of view from that was to double a point of
0:14:39our our that it's
0:14:41much more what was as compare
0:14:44at least is uh a non of a step size
0:14:48yeah we use a a simple get about that that
0:14:51a vector just two
0:14:53the adaptation
0:14:55double talk
0:14:59i in it is a a a a few conclusion
0:15:01so we propose
0:15:02this out that point text of for echo cancellation
0:15:07we can take advantage from a its computation of complex
0:15:12which is low as from that is a class
0:15:16we think that is important that the value of a step side
0:15:20somehow somehow not about made
0:15:22this means that we do not use any additional power uh
0:15:27i'm sure one by one
0:15:31we also note this is that
0:15:34these more what used to
0:15:36uh near end signal evaluation
0:15:38and most important to double
0:15:42thank you very much earlier
0:15:51for that
0:15:52well we do you have some time for
0:15:53i if anybody like
0:15:58using again
0:16:03presentation presentation just them curious about two
0:16:05the step size it becomes negative when the values of the
0:16:08design a value is a lot to than that of the there
0:16:17in this problem
0:16:19this is a very nice for
0:16:21because it we ski but it to meet some practical issues yeah
0:16:26it is a recommended fast
0:16:28to at
0:16:29a small positive constant yeah
0:16:31just to avoid you bytes
0:16:33at most important
0:16:34we should they is the absolute but
0:16:37this a three star is then
0:16:40but but what i
0:16:42too much
0:16:51inking your simulations used in order
0:16:53a projection order
0:16:54for for for yes
0:16:56and is the method sense to the project no
0:16:59no we just to choose this value
0:17:02it's a for complex
0:17:04we try was my
0:17:06projection of
0:17:08you have
0:17:09and in your a personal experience
0:17:11what projection order do you find is really the the right
0:17:14and it to be working in
0:17:16for for
0:17:19i think that uh to three we well we should not to take a larger than eight
0:17:26but also because of the complexity issue you by that you for this these some uh
0:17:32that's a diffusion map for completely because in a fine as the fine projection of what you
0:17:37so main problem is how to invest
0:17:40the out
0:17:41i i have a project the higher projection of but we been blind or if you a quotation my
0:17:46recently the as far as i
0:17:48no was that are some more efficient manner
0:17:50for example
0:17:52you you all
0:17:53because the models
0:17:56or the might be end
0:17:57variation of by
0:17:59well also a
0:18:01which she uh if you lead that for this operations so we can even go higher
0:18:05a project
0:18:09and other question
0:18:15i think i have the opportunity to last one um
0:18:18a more general question concerning the difference between acoustic echo cancellation
0:18:23and network
0:18:26could you clarify which
0:18:28i think you first results at least as in the network
0:18:32a the was the
0:18:34so is the star
0:18:35a point of
0:18:36two and eight type of that if you that
0:18:39for a network that for consideration and respect to
0:18:43that form but that's
0:18:44send send you because of this i
0:18:47of the
0:18:47because i buy it can fall you by room acoustic a more cancellation if we properly choose
0:18:54the proposed to eight five
0:18:55this was that he's and defines the chose of the I the nlms out
0:18:59which is
0:19:00somehow of a host to this
0:19:04part think the
0:19:05of course the
0:19:06we can use for example a low as far as and uh as i
0:19:10we proposed a few
0:19:12a one year ago goes as the sparseness control coding which can
0:19:16a beep on uh
0:19:18well efficient as compared to ipnlms mess if we know as this from miss what it's to make some the
0:19:25so we is important to choose the proposed
0:19:30well that's great thank you very much and