Přepis řeči - ROBUST AND LOW-COST CASCADED NON-LINEAR ACOUSTIC ECHO CANCELLATION

0:00:15	you
0:00:16	and
0:00:16	right
0:00:17	my name is
0:00:19	though
0:00:20	so more people
0:00:21	or you know
0:00:22	action
0:00:24	a a can go you know that your account for a colour version of
0:00:27	yeah
0:00:28	more work competition
0:00:32	i'm i'm so through a uh
0:00:34	a yeah
0:00:35	it
0:00:37	i enough to
0:00:38	it'll "'cause" using a that's got model
0:00:40	no you know it
0:00:43	are first and to reduce as the acoustic echo or then i will present
0:00:47	they're model be used for acoustic bass
0:00:49	and then using this model out was on the on know to stick
0:00:53	it can can all that we proposed in this paper
0:00:56	and then uh was on the test results on the uh i would and my presentation for compression
0:01:03	so
0:01:04	yeah
0:01:05	it will
0:01:06	the are core problem i as one that far and speech your signal is given by the microphone
0:01:12	we have a a a cool white white and you
0:01:14	and then sent send back to the a for in speech your and this was did you ain't in that's
0:01:18	what we do and to we so
0:01:20	someone
0:01:21	but to mission in the network
0:01:23	to solve this for and approach used use the acoustic know
0:01:27	acoustic echo can
0:01:28	that try to estimate it was not once press
0:01:31	in the microphone signal and the where and what the more while is close to white had is close to
0:01:36	what
0:01:36	we will this the if effect off for a cool
0:01:40	two
0:01:41	estimate this it would i i C need to
0:01:44	and estimate of five the acoustic channel and uh
0:01:48	the most common
0:01:49	approach to use
0:01:51	is data a you know or environment J where we was that it was channel
0:01:55	the it energy or is a convolution between the
0:01:57	far and signal
0:01:59	and they
0:02:00	acoustic of of you work with than that of
0:02:02	of the down the not speak or the acoustic channel
0:02:05	and the microphone a
0:02:08	but uh are to the a some by phone arm reach a right on that we use a load device
0:02:14	a devices so would be non you know is
0:02:17	some
0:02:17	devices and to be some non not is that the all speaker
0:02:21	and uh these uh are non no just mentioned on told nonlinear environments
0:02:25	and here we can see where some simulation
0:02:27	well i and would be so
0:02:29	in that all speech your
0:02:31	some known note
0:02:32	a pro three non know
0:02:34	you are it's so the output or the speech or that we suppose that they those be drawn to be
0:02:39	some known not
0:02:40	and we can see how
0:02:41	when D
0:02:42	none no it's if it in how they
0:02:45	different you know a a good soon
0:02:46	performance trees
0:02:48	um sort is for and you know
0:02:50	we use a a a nonlinear adaptive filter
0:02:53	the not not not that if you choice is a system
0:02:55	that try to estimate did you know what on and non you know cost
0:02:59	of the
0:03:01	acoustic channel
0:03:02	and uh i
0:03:03	try to estimate the it they only or equal on suppress it from day
0:03:08	it was signal
0:03:10	from the microphone signal
0:03:11	but this problem
0:03:13	this a
0:03:14	more to and thirty some form a as we can see your we use many adaptive filter
0:03:19	this will introduce some slow convergence
0:03:21	and also as and they do not case
0:03:23	the our system depend on them on that we use
0:03:26	if well we use a a wrong model or in the in on you know
0:03:29	in our on your is
0:03:31	we not to improve the performance so we need first
0:03:33	a a
0:03:34	mortar are often only you is in this case me
0:03:37	use a model of the house be john we for that
0:03:40	a of the known not just come from that speaker
0:03:44	in our approach we use a a a a a of an unknown to
0:03:48	this is a
0:03:49	this come from the does that you have also
0:03:52	from some well for no now speak your on a be have seen that this
0:03:56	more that is uh
0:03:58	is are interest interesting to model or the not speech or no it's if it
0:04:02	and we and also assume that uh in this case
0:04:05	they now speak your
0:04:07	features
0:04:08	um
0:04:08	so a very small compared to the acoustic channel your we model
0:04:13	and you know stick channel which is a
0:04:15	and acoustic uh the acoustic channel or and the microphone off
0:04:19	so we can suppose that uh
0:04:22	as the acoustic channel you we can suppose that the are stick channel is a you know a filter on
0:04:27	also the microphone is the you know a two so would it to that's got of the two channel
0:04:31	can be
0:04:32	more than a city in from don't we far
0:04:34	the are you see
0:04:35	of the acoustic channel we suppose that that six channel is a hard
0:04:39	variable so we can assume that
0:04:41	this part is highly variable
0:04:43	and this part is a a less variable
0:04:45	but uh uh is that
0:04:47	and a is a nonlinear
0:04:51	in our proposed up what our aim is to
0:04:55	to use it that's got a model of a a a at that's got model
0:04:58	where we try to first estimate the output of the non-speech you're
0:05:02	sure and then we use a you know adaptive filter
0:05:05	estimate it you know of of the last minute that's got of the
0:05:08	acoustic channel and the my proof
0:05:11	where we use this mortar really gets marked improvement because in general
0:05:16	they've i'd every you of eigen the channel the whatever
0:05:19	of the acoustic path
0:05:20	come from the acoustic channel so
0:05:22	that's mean we
0:05:24	that's me we only need to wind up a steep of change only need to re that the A C
0:05:30	and we don't need to
0:05:31	we i of the people so
0:05:35	you are so we make a compression between the
0:05:39	i a model of are the power power of to a power ch on they that's got is more than
0:05:44	that to be
0:05:45	was true
0:05:46	yeah we can see that there are much much but if you and that's mean
0:05:49	the
0:05:49	first so channel here
0:05:51	is that you and to the convolution that unit it chan
0:05:54	and difference channel in the mike of in the proposed so
0:05:58	but uh are what we can expect that this just one we use
0:06:03	a got a model that we have a high or
0:06:06	means is a higher minimum error or to day
0:06:10	but model because in the power model
0:06:12	as we
0:06:13	we estimate only one channel
0:06:15	we have a there are about to show that
0:06:17	we can have also a minimum error here your and you know that your
0:06:21	a are in the proposed still so
0:06:23	the the one minimum or you higher compression
0:06:26	power power to
0:06:27	probably case
0:06:30	and uh
0:06:31	where what's something that we can uh
0:06:34	we
0:06:35	something that's
0:06:36	we can see that
0:06:37	if the H change here
0:06:39	we need to race
0:06:40	here estimates
0:06:41	or different job here and in this case we just need to estimate
0:06:45	only did you know a lot of the feature
0:06:49	or we will go to how we estimate our model
0:06:52	we just we first given even the different expression of well
0:06:55	oh signal
0:06:56	you are white P
0:06:58	is supposed to be
0:06:59	is the
0:07:00	what is the output of the people sister
0:07:03	and and white N is the
0:07:05	why and
0:07:06	is that
0:07:07	it was no
0:07:08	and the estimate it also use the same we have
0:07:10	white we to with put of the proposed on the sum of
0:07:14	they white hats
0:07:15	we divide the
0:07:17	what we for the H
0:07:19	we give our they estimated
0:07:21	it was your
0:07:22	oh and that or is the difference between the estimates
0:07:25	the different in that it was not the estimate of
0:07:28	a and it is method it was you know
0:07:31	you are we use a image for estimation to see how
0:07:35	all of it over here yeah
0:07:36	in a me
0:07:37	in a
0:07:38	test
0:07:39	yeah know we estimate our the output
0:07:42	the out of here
0:07:43	can be
0:07:44	estimated it using the mean square error
0:07:47	by day
0:07:48	cross correlation between
0:07:49	the
0:07:50	why
0:07:51	why i which supposed to be day
0:07:54	it was not and they're
0:07:55	white white you which is that out speech you're out
0:07:58	but in a red we don't have a just to this
0:08:01	white P signal
0:08:02	you G is the output of a a speaker
0:08:06	so this you ask the they are that
0:08:08	if i day
0:08:09	why
0:08:10	the output of the that speaker or is high it
0:08:13	we will decrease the common knows rate of for
0:08:15	oh you know a feature
0:08:18	and then we go to the estimated of the people still so to and this is
0:08:22	what one call
0:08:23	so for choices a
0:08:25	for don't in the people so
0:08:27	so we can see that
0:08:28	it jar
0:08:29	so for
0:08:31	a it's of the joke can be estimated
0:08:33	a a like
0:08:34	cross correlation
0:08:35	and the what to correlation and and the adverse or to coalition
0:08:39	of
0:08:39	the output
0:08:40	for each people so
0:08:42	this a
0:08:43	we the cross correlation between
0:08:45	the output of one
0:08:47	of the
0:08:49	P you have uh
0:08:50	so for job and the output of the
0:08:53	or or or something to
0:08:54	and uh that's um
0:08:56	that's why an order be but one
0:08:58	a power that system is used
0:09:00	people proposed to
0:09:01	and not to go in addition
0:09:02	this
0:09:03	if it
0:09:03	because one day
0:09:04	system of the one i
0:09:06	this uh a really put to zero
0:09:10	so
0:09:11	after that i we're
0:09:12	sure how we added C estimate
0:09:15	the different should to this is the
0:09:17	and the normal and uh
0:09:19	the need
0:09:20	you know me
0:09:22	this to mean score approach we use your
0:09:24	we can see that uh if
0:09:25	it's uh and M S words
0:09:27	on that we can see that in each case yeah we used
0:09:30	the estimated of of the wood of or was so
0:09:33	on and also for the
0:09:34	people to so for some for joe
0:09:36	we use the estimates
0:09:37	of the know
0:09:38	uh use
0:09:39	as we use the
0:09:41	estimate of the you noise that's mean we need to
0:09:44	ooh some global step size in the put
0:09:46	so estimation
0:09:47	two and of all you for what these that we just
0:09:50	and and also as most of a
0:09:53	a a porsche
0:09:54	people in in said i the or
0:09:56	filter or two zero you in this case in that's good of course
0:09:59	we cannot not this anything goes
0:10:00	we we not change a remote uh
0:10:03	change
0:10:04	change the effect of our of it so we need to use it as one of the tapes
0:10:07	in the proposed still
0:10:09	you equal to one
0:10:11	now would go to
0:10:13	oh test result
0:10:14	well we first give day
0:10:16	oh to sit up know we use
0:10:18	five people so that's mean the P
0:10:21	sure is equal to
0:10:22	five
0:10:22	and and for each people cells
0:10:25	so it for yourself a job we use
0:10:27	one hundred taps
0:10:29	and and for you know job
0:10:31	we use two hundred
0:10:32	and in a lot of to be used three
0:10:34	hundred taps
0:10:36	we use a so
0:10:37	signal to noise or racial
0:10:39	thirteen db V on fourteen db that is a
0:10:42	it was to to denoise noise or which
0:10:46	so the first one this
0:10:47	when we and
0:10:48	the first was your
0:10:51	so the different case where a some suppose that
0:10:54	we
0:10:54	how some they echo path changes
0:10:56	so we can see that for different point of the it of changes
0:11:00	the different should do we had different
0:11:02	know the power of it to is the what's
0:11:04	we is the power model
0:11:06	and the proposed it
0:11:07	yep was an more your
0:11:09	use a three
0:11:10	taps
0:11:11	in the people so on five taps and the
0:11:14	people so and uh i think we one is the normal one a a mesa
0:11:18	and then the ms so i
0:11:20	one we went you in the change parts
0:11:23	we can see that when the first echo path change all eyes
0:11:26	the and lms on
0:11:28	on the power model
0:11:29	the part and model
0:11:31	do you better result convert
0:11:32	the propose it model this is due to the fact that
0:11:35	as the first change
0:11:36	at the first change
0:11:38	pop and model people still a tough
0:11:40	a good uh
0:11:42	i and i curse estimate of the people still
0:11:44	so feature
0:11:46	and a then we can see that at this point to a proposed approach
0:11:50	you better results and to the second a couple of change
0:11:53	we you have a terms that for that was model
0:11:56	about to do or more
0:11:59	the a second uh this so is when we change
0:12:03	we change D
0:12:04	it group but they
0:12:06	delay in the echo path
0:12:07	that's mean we
0:12:08	suppose that the
0:12:10	we and to be some face is in the a couple of
0:12:13	and as for the previous case we can see that
0:12:16	when first first date changes in
0:12:18	we have better performance for a power of two
0:12:21	and the and ms that to a much for sir
0:12:24	then they propose model try to calm
0:12:27	try to we i that and we have better a performance
0:12:30	and
0:12:31	when we go to the next
0:12:33	a cool
0:12:33	do date change changes we
0:12:35	the
0:12:35	what wasn't model
0:12:37	have a better convergence than the all of them but the power tool and the
0:12:42	you know and and ms
0:12:44	and
0:12:46	the and or or to go to
0:12:47	conclusion
0:12:48	uh we for sure and that's got up was to D in the acoustic
0:12:53	the cool constellation we have shown that this uh
0:12:56	more than is more robust
0:12:58	for the
0:12:58	a group of changes and need to get change
0:13:01	a if a C a to soup of this work is to reduce the complexity of the system a small
0:13:06	so
0:13:07	one you know system are really complex
0:13:09	on also to have a a bit top one for one to people so
0:13:12	because a
0:13:13	when did echo path change
0:13:15	we should do we used the people still
0:13:18	adaptation and the uh
0:13:19	it's a improve day
0:13:21	it's wouldn't programs
0:13:24	and you for attention
0:13:32	i you we can takes
0:13:33	question
0:13:34	a
0:13:50	that that's good was like a you
0:13:53	yeah it's can
0:13:54	it's can i
0:13:55	generate can give you as a us on
0:13:57	local or local minima
0:13:59	but so what we assume your that's
0:14:01	the acoustic but
0:14:03	is not
0:14:04	a it's up
0:14:04	but or not that's that's on the elements
0:14:06	so we don't have very dot
0:14:09	don't face
0:14:10	this kind of problem
0:14:11	or they are what they local alone
0:14:14	but what you know as one we was that the
0:14:17	it group of these are we likely to be a
0:14:20	step
0:14:31	B
0:14:32	they they what we model
0:14:35	as this uh they do not part of the echo path change
0:14:39	if we was that this case
0:14:40	so we not to have a a the problem
0:14:44	a for a local you one one
0:14:47	a
0:14:59	yeah it's closer this
0:15:01	a a a a whole to proposed so
0:15:12	take
0:15:16	yeah yeah yeah
0:15:18	get it
0:15:19	it's pretty however
0:15:20	yeah
0:15:22	yeah how we initialize ours
0:15:24	a but uh as i say we need to
0:15:26	was on or step size
0:15:28	in the eight people so
0:15:30	to avoid the uh
0:15:31	for
0:15:33	that
0:15:34	as we need to use the estimate of the you know to
0:15:37	in the estimation
0:15:38	so
0:15:45	our project is set i to see what we need to
0:15:48	what one of the
0:15:49	one of it that to
0:15:50	equal to one
0:15:52	and
0:15:55	and yeah
0:16:15	they you
0:16:16	a group of can be more
0:16:26	i to what can be more
0:16:28	but
0:16:35	H one had and yeah it's a yeah it's can be water
0:16:40	there
0:16:41	yeah that's what's up in this case that's
0:16:43	if i H one had is not
0:16:46	we equal to be H
0:16:48	the
0:16:48	who wish your
0:16:50	it's will be in the a you know a lot of H
0:16:54	and the and verse with a or of that that's what
0:16:58	it's not
0:17:10	yeah it's
0:17:11	it's a very different close
0:17:13	if for suppose that it's hot
0:17:15	i H one hot
0:17:17	and H two hot
0:17:18	a a close to each other if a if numbers
0:17:21	you
0:17:21	this of a job you close to
0:17:24	a yeah
0:17:26	yeah sorts a good don't
0:17:28	i do make that this but
0:17:29	it will not
0:17:30	i
0:17:31	Q
0:17:35	but to to you
0:17:37	can is in low so
0:17:39	they reported
0:17:40	yeah
0:17:43	constraints
0:17:44	H one
0:17:47	any other as two
0:17:51	yeah just
0:17:52	one small question myself yeah
0:17:53	i i is to press
0:17:54	and and like a three it and it is that is using the cell phone in it
0:18:01	in
0:18:02	it's like three
0:18:03	yeah yeah
0:18:08	uh so that is
0:18:10	we use
0:18:10	yeah
0:18:11	so yeah you
0:18:13	uh
0:18:13	i as is a uh that's good okay
0:18:16	and but the most serious question is
0:18:18	and
0:18:19	i mean try to understand
0:18:22	um well
0:18:23	a cool and you what is the magnitude
0:18:26	cycle i
0:18:27	and the thing to note a nonlinear
0:18:30	signal part
0:18:32	i i want to let that the data which are processed here was from a real set or whether it
0:18:37	was seen the said that
0:18:39	if we see any improvement let's say between zero and i think about four db
0:18:43	yeah a really
0:18:45	and
0:18:45	is that's indicative of what we might
0:18:48	i
0:18:48	yeah in a real
0:18:49	uh my about an application
0:18:53	for L C for real application this so the our is it is form solution
0:18:58	but they also need to read the it did you not a couple of used
0:19:01	come from a you know or or a a a a a you this so
0:19:04	or not this estimate
0:19:06	but uh i in real just what we have seen it's to we use in general are it's some sense
0:19:11	signal
0:19:11	we can see that there is a
0:19:13	reorder out T we have a a a a life it's
0:19:16	of you know you know is
0:19:17	and i
0:19:19	i fifty for it's
0:19:20	and to seem to focus on this kind of on this problem
0:19:23	okay is that if
0:19:25	is are the real test set which you use pos it from a um
0:19:28	a um
0:19:29	a a stick all
0:19:31	i
0:19:32	it's a far set it's a cell phone
0:19:39	so that's like the speaker

ROBUST AND LOW-COST CASCADED NON-LINEAR ACOUSTIC ECHO CANCELLATION

Echo Cancellation

Přednášející: Moctar Mossi Idrissa, Autoři: Moctar Mossi Idrissa, Christelle Yemdji, Nicholas Evans, EURECOM, France; Christophe Beaugeant, Philippe Degry, Infineon, France