Přepis řeči - SPATIO-TEMPORAL SIGNAL PREPROCESSING FOR MULTICHANNEL ACOUSTIC ECHO CANCELLATION

0:00:13	um sort of the delay a we come to my presentation on a spatial temporal signal
0:00:19	was
0:00:20	multichannel acoustic a cook
0:00:22	relation
0:00:23	uh that's
0:00:24	or like like to give an outline of my presentation
0:00:29	i would introduce my presentation by you
0:00:32	brightly
0:00:33	maybe on the problem of multichannel acoustic echo it in constellation
0:00:38	a a a a a a that a a approach
0:00:40	i i i would present
0:00:42	yeah
0:00:43	in my lecture
0:00:44	can say
0:00:45	uh
0:00:46	massive multichannel reproduction systems and for that i would like to
0:00:51	we interviews
0:00:52	some basic properties of a method of the channel
0:00:56	reproduction systems
0:00:58	a that i one
0:00:59	introduced a the novel approach for spatial can problem
0:01:03	preprocessing
0:01:04	and
0:01:05	give
0:01:06	a to or
0:01:07	uh um
0:01:08	a
0:01:09	a roles of this i
0:01:13	or practical
0:01:14	implementations and after that i
0:01:16	to
0:01:17	with no
0:01:19	so i
0:01:20	think
0:01:21	you have a
0:01:21	seen
0:01:23	similar a
0:01:25	block yet from
0:01:26	in this section so forth that
0:01:29	i
0:01:31	i is suppose you know what is says you know so we have here just a a a a a
0:01:35	and and and you and strong
0:01:38	and i i yeah
0:01:40	all what acoustic E
0:01:42	constellation it to create
0:01:45	a replica of sandy the or in room
0:01:49	um
0:01:51	model model as um my
0:01:54	F I R
0:01:55	filter
0:01:56	and the that
0:01:58	got
0:01:59	of
0:01:59	that my my by a filter
0:02:01	okay
0:02:02	are usually a minute
0:02:04	by by at least squares optimized addition
0:02:09	the optimal
0:02:10	solution
0:02:12	right
0:02:13	a or is given by the you know how to equation
0:02:17	also known
0:02:18	and a normal
0:02:20	equation here
0:02:22	we have that in terms of
0:02:24	are are X
0:02:25	it's which uh
0:02:27	if uh inverse of the
0:02:29	correlation metrics of the loudspeaker the signals in then used and
0:02:33	role
0:02:34	and are it's Y D notes
0:02:37	that
0:02:38	a cross correlation
0:02:39	metrics tricks between the loudspeaker signals
0:02:42	and the microphone signals
0:02:44	he's any and room
0:02:47	um
0:02:49	in
0:02:51	in practise
0:02:51	there
0:02:53	at the filters
0:02:55	are
0:02:56	the term and
0:02:57	at optimally
0:02:58	because
0:02:59	that acoustic of the own
0:03:01	could be
0:03:02	i of over at a time
0:03:05	is that it can be
0:03:06	shown
0:03:07	that
0:03:08	that's speed
0:03:09	of con veterans
0:03:11	oh adaptive filters depends of on that i gains
0:03:15	spread
0:03:16	of the autocorrelation metrics
0:03:19	and for some fleeting
0:03:21	in a typical
0:03:22	uh a for the the like
0:03:26	communication
0:03:27	scenarios we have only
0:03:30	you
0:03:31	if source source as in the are end room
0:03:34	and Z sources
0:03:35	are
0:03:36	right by
0:03:38	many loudspeakers speakers in there
0:03:40	near and room
0:03:41	hence we have a
0:03:43	and
0:03:44	ill conditioned
0:03:45	are
0:03:46	X six
0:03:48	and
0:03:50	how to cool
0:03:51	where
0:03:52	just problem
0:03:53	that i do yeah
0:03:57	is to preprocessed process that allows loudspeaker signal
0:04:00	in that near in troll in order to
0:04:03	decorrelate the channels with rest
0:04:05	but to each other
0:04:08	that
0:04:09	the preprocessing techniques
0:04:12	have to fulfil fill
0:04:16	at least
0:04:16	in three but
0:04:19	and first one is
0:04:20	for for of that convergence enhancement
0:04:23	that's
0:04:23	second requirement
0:04:25	if
0:04:25	is that a preprocessing technique
0:04:28	that doesn't introduce
0:04:30	all people
0:04:31	um uh the distortion on the signal
0:04:35	and
0:04:36	also
0:04:37	that
0:04:38	there
0:04:38	preprocessing technique doesn't
0:04:40	in use
0:04:42	in the yeah
0:04:44	uh
0:04:44	compute uh
0:04:46	uh
0:04:47	audition them
0:04:48	complexity complexity factors
0:04:51	for
0:04:53	to oh
0:04:54	for us
0:04:54	T your isn't to their production systems
0:04:58	you can find
0:04:59	many approaches
0:05:01	for preprocessor loudspeaker signals a
0:05:04	and T i mention only two
0:05:07	the first one was introduced by
0:05:09	the the is T
0:05:11	and to the idea
0:05:14	oh it is to rate up to introduce nonlinearities as the signal
0:05:19	this approach
0:05:20	feature a very low complexity
0:05:24	but that distortion that it
0:05:26	in uses
0:05:27	bit come quite to a problem for high quality
0:05:30	signals
0:05:31	for for them for music
0:05:33	conference or something like that
0:05:36	and uh and a second to a perceptually when motive that up holds well
0:05:42	introduced used by you can hear to
0:05:46	and the i
0:05:48	B is this
0:05:48	one have
0:05:50	for use on a frequency selective
0:05:52	phase modulation
0:05:54	of the a loudspeaker
0:05:57	signal
0:05:59	so the uh he of of that a loudspeaker signals
0:06:03	are decomposed
0:06:04	in in bands
0:06:06	in the
0:06:07	by means of uh analysis filter
0:06:10	and the sub-band signals
0:06:12	uh a or the face of the subband signals
0:06:15	and modified
0:06:17	in a frequency dependent
0:06:20	am modulating signal
0:06:25	and the now i
0:06:28	a to come to the second point
0:06:30	and what is like to introduce
0:06:33	that massive multichannel reproduction systems
0:06:36	the sound field
0:06:38	sinc to is this
0:06:40	systems
0:06:42	in at is that to use this
0:06:45	of the sound field to put used might have the actual source
0:06:49	S
0:06:51	with an an extended yeah yeah
0:06:53	V
0:06:54	using and i'm bill
0:06:56	of a speakers on the boundary of that area
0:07:00	here
0:07:01	the
0:07:01	the delta be in
0:07:04	that go
0:07:06	this many techniques
0:07:07	um you for
0:07:10	solving this
0:07:11	problem
0:07:12	and the
0:07:13	we
0:07:14	some of them are are this specialised
0:07:17	on specific
0:07:19	a a is you for example is that near field compensated tie i what i'm based on it
0:07:26	which is the
0:07:27	a specialised
0:07:28	on a circle and say and
0:07:31	and is
0:07:32	we have also for a planet and lean yeah and i is
0:07:35	is that
0:07:36	so called
0:07:37	it's spectral division method
0:07:40	it's the M
0:07:42	and the
0:07:43	more or less
0:07:45	is more one
0:07:49	uh
0:07:49	and i rate in an independent
0:07:52	technique technique it is is a a a a a W F S is the wave field synthesis techniques
0:08:00	which words for linear plan a and other convicts
0:08:03	and i image
0:08:07	for sure as a a low or
0:08:10	to kill implementations of for sound field sent to just
0:08:14	are
0:08:14	discrete
0:08:16	and therefore
0:08:19	the except
0:08:20	the the except if something like an aliasing frequency
0:08:24	and we can say that the control ability of the sound field is available only up to a given
0:08:31	frequency
0:08:34	and
0:08:35	here we see use this simulation one and linear very
0:08:38	with
0:08:39	twenty loudspeakers speaker
0:08:41	and here at the frequency have it
0:08:43	it's hundred
0:08:44	have which is below the
0:08:47	and you in frequency of of this yeah
0:08:50	right
0:08:51	that wave front can
0:08:53	be response
0:08:54	set correctly
0:08:55	and here for the two thousand
0:08:57	a time the test which i
0:09:00	all the figure that L in frequency we see that
0:09:03	that
0:09:04	where front
0:09:04	and a be reconstructed
0:09:07	firstly
0:09:08	so
0:09:09	and the
0:09:10	two our
0:09:11	problem
0:09:12	that we like to implement
0:09:14	and uh
0:09:15	uh acoustic echo canceller
0:09:18	for such a set up who with with massive multichannel
0:09:21	systems
0:09:23	this problem is
0:09:24	a challenging
0:09:25	not only be a is that is the autocorrelation matrix as
0:09:30	are are very last and in condition
0:09:32	when
0:09:33	addition
0:09:35	uh it can be shown that the someone
0:09:38	you to synthesized
0:09:40	by a massive multichannel
0:09:42	the production techniques
0:09:44	are
0:09:44	very sensitive to any day deviation
0:09:48	on the tiring
0:09:50	signals
0:09:52	uh
0:09:53	do you to build a preprocessing
0:09:57	and to hear it should be
0:09:59	not
0:10:00	that is that best section
0:10:03	or from from and uh so i got the point of view
0:10:06	that's a perception of
0:10:09	that is synthesized
0:10:11	wave fields
0:10:12	by
0:10:14	sound and to the synthesis
0:10:16	techniques
0:10:17	are
0:10:18	completely
0:10:20	i is completed different the then of stereophonic funny
0:10:24	the based techniques
0:10:26	and haynes
0:10:28	any
0:10:30	temporal preprocessing should be avoided
0:10:34	the so we can
0:10:35	do and a preprocessing
0:10:37	techniques
0:10:38	and how to cope with this problem
0:10:41	and
0:10:42	therefore we
0:10:44	have
0:10:45	it a novel approach for spatial or temporal i E C preprocessing
0:10:50	and this is is never a novel approach
0:10:53	basis
0:10:54	actually on how a very
0:10:56	all
0:10:57	with them
0:10:58	namely prevention is better than Q
0:11:02	and
0:11:03	what we would like to do
0:11:05	what we do
0:11:06	would like to do if actually
0:11:09	two
0:11:09	that
0:11:10	but
0:11:11	here and in addition and that the channel
0:11:15	and uh processing you
0:11:20	it
0:11:21	is that a i is that a loudspeaker signals in such a way
0:11:25	that is that a wave feel is set to zero euro the positions
0:11:29	of the i
0:11:30	ones
0:11:32	so we light
0:11:34	two
0:11:35	D lee
0:11:36	is are listening yeah yeah in two
0:11:39	two the ones
0:11:40	one own is quite where some microphone and i is supposed to be
0:11:44	and then another listening
0:11:46	yeah yeah
0:11:47	where is the desired
0:11:49	feel is a construct a
0:11:51	lee
0:11:53	is the
0:11:54	but
0:11:55	problem of creating
0:11:57	own on i i'd this not new and the in the literature
0:12:02	you can find many
0:12:04	start days
0:12:05	on it
0:12:06	and here i have listed
0:12:08	only some also
0:12:11	but to
0:12:13	here i i'd like to introduce
0:12:15	and you and a local
0:12:17	approach
0:12:18	for a
0:12:19	at uh the uh
0:12:21	creating zones of quiet with linear
0:12:24	and i this approach is best
0:12:27	it just straight
0:12:28	by
0:12:29	uh uh but is illustrated but be thing on the method of spectral division method and therefore i would like
0:12:36	two
0:12:37	just
0:12:38	give that say or tickle be six
0:12:40	of
0:12:41	that
0:12:41	mm method
0:12:43	and in is it is this method
0:12:47	assume assume what
0:12:48	or or to be is is actually
0:12:50	one the idea that some wave
0:12:53	feel
0:12:54	reproduced produced by a linear distribution
0:12:58	of the second outing
0:12:59	source of
0:13:01	can
0:13:03	B
0:13:04	it it's press
0:13:06	by a convolution of the thinks signals of
0:13:10	is that a second so that's was the green
0:13:13	function
0:13:14	of that services
0:13:15	and that that are green
0:13:17	functions can be and just to
0:13:19	as a
0:13:20	spatial temporal transform domain
0:13:22	to but also a
0:13:24	secondary source
0:13:27	in the following derivation we will restrict our configuration on a reference line
0:13:33	and the
0:13:34	uh what if we do that we can
0:13:38	see
0:13:39	is that
0:13:40	we have a
0:13:41	one time in the novel
0:13:43	the um
0:13:45	the one dimensional convolution
0:13:48	and for it
0:13:49	exact convolution theorem holds
0:13:51	so if we do a fourier yeah transform
0:13:54	we get a multiplication
0:13:56	and to by to be an announcement we get an expression for that mean
0:14:00	function
0:14:01	for
0:14:04	uh
0:14:04	so and uh
0:14:05	here is idea for creating the role of quite is actually just by multiplying
0:14:10	that desired wave field white i by L we don't by a specific we
0:14:15	and the if
0:14:17	we we be formally that in the wave number can mean we get them a
0:14:22	and expression for the desired wave field
0:14:25	in the wave number mean
0:14:27	and we get after that an expression for the subscribing
0:14:30	signals
0:14:31	that
0:14:32	to to prove that our approach
0:14:34	can i is applicable
0:14:36	to a a acoustic it constellation we have computed that attenuation innovations that can be done
0:14:42	by
0:14:43	this approach by creating as a of quite and we uh and
0:14:48	put
0:14:49	here
0:14:50	and microphone
0:14:51	and
0:14:52	right that's in an we shouldn't with that
0:14:54	a point which
0:14:55	is supposed to be
0:14:57	in the sweet spot
0:14:58	and we got an at wishing up to
0:15:01	seventy
0:15:02	D V
0:15:03	which is comparable was to become
0:15:06	yeah it cool return last
0:15:08	enhancement quite
0:15:11	S
0:15:12	i have a
0:15:13	shown
0:15:14	hmmm is that uh the control ability of there
0:15:18	wave if it's can be done only up "'cause" the anything frequent and frequency and they for for higher frequencies
0:15:24	acoustic it go
0:15:25	constellation or a typical acoustic echo equal constellation
0:15:28	must be still
0:15:30	a applied
0:15:32	and therefore for
0:15:33	we propose
0:15:34	here
0:15:35	at
0:15:35	divide in
0:15:37	the or or a yeah i don't like or decomposing "'cause" the signal into to sub bands and for a
0:15:43	lower
0:15:44	frequencies
0:15:45	just to so
0:15:46	right to use our approach and for i higher frequencies
0:15:50	to apply and the acoustic you constellation
0:15:53	yeah and say that approach
0:15:55	can be also
0:15:56	uh
0:15:57	use
0:15:58	well with the equal colour
0:16:00	and is
0:16:01	which are
0:16:02	more interest
0:16:04	interesting
0:16:05	for spatial audio reproduction systems
0:16:08	for example by applying
0:16:10	that output
0:16:12	but it represented presented by a problem
0:16:16	and i come to the conclusion
0:16:18	and the intuition of my work well
0:16:21	to show that is that
0:16:23	at to the acoustic you constellation can be done in a a distributed
0:16:28	manner
0:16:29	on the loudspeaker a
0:16:31	side
0:16:32	and on some microphone side and not
0:16:34	only
0:16:35	at it has been
0:16:37	the done
0:16:39	does on the microphones side
0:16:41	and i have shown one yeah
0:16:43	um a method to can for creating zones of quiet as linear allows loudspeaker i
0:16:48	and is i have sons that
0:16:50	that imitation that we have a
0:16:52	yeah is
0:16:53	and and on that
0:16:54	uh anything frequency
0:16:56	of the loudspeaker a rate
0:16:58	therefore for frequency selective implementation should be done
0:17:01	thank
0:17:09	uh we have time for questions two
0:17:16	T
0:17:17	uh i would you like
0:17:20	i got got a scalar
0:17:28	oh
0:17:29	i was wondering if
0:17:30	you have
0:17:31	oh
0:17:32	people
0:17:34	in side
0:17:35	you're or right
0:17:36	would they change
0:17:37	uh
0:17:38	as they move around
0:17:40	with you
0:17:41	the quiet
0:17:42	the G
0:17:45	yeah actually
0:17:46	we
0:17:48	yeah
0:17:49	we have
0:17:50	um
0:17:51	um
0:17:53	consider
0:17:54	only
0:17:55	that analytical cave yeah and we have a set for the right functions a free feel
0:18:01	yeah
0:18:02	and
0:18:03	hmmm
0:18:04	as the for the free field the green function
0:18:07	but for
0:18:08	a practical of imitation should be implemented in now
0:18:10	at that
0:18:11	man
0:18:12	and
0:18:14	so so we will have a just a mismatch between that's that's supposed
0:18:18	green function and a real one green function
0:18:21	and therefore we will not get
0:18:23	is that
0:18:24	performance of
0:18:27	going to quite on as we have
0:18:30	my my
0:18:31	uh what would the geometry of a relationship of the microphone
0:18:36	and a loudspeaker
0:18:38	would like thing to the lead the microphones
0:18:40	you're the low
0:18:41	result
0:18:42	just
0:18:42	in
0:18:43	you know maybe you
0:18:45	consider a circle
0:18:47	uh
0:18:48	you know where have you thought about that would be a very sad
0:18:51	or
0:18:53	a microphone geometry we we are independent actually from the microphone german trace
0:18:58	a a always a loudspeaker to and we have
0:19:00	simulated with omnidirectional
0:19:02	lower speed
0:19:05	oh
0:19:05	okay
0:19:07	and you did you in
0:19:09	room reflections so
0:19:11	so i just want
0:19:12	is the low frequency stuff you you do by
0:19:15	well making the by
0:19:17	yeah
0:19:18	it becomes an active noise can
0:19:19	a problem if you
0:19:21	microphones phones in the five
0:19:23	you can use that
0:19:25	able
0:19:27	to full the null
0:19:28	at a point to the mic
0:19:30	and an active noise
0:19:31	relation
0:19:32	yeah maybe
0:19:35	it's is a good idea i now we have a we haven't considered that
0:19:38	it
0:19:39	seems to be
0:19:40	with a
0:19:44	i
0:19:48	yeah think is that when you are no you the microphone
0:19:53	was it microphone
0:19:54	and the the markers
0:19:55	yeah
0:19:56	that will not
0:19:58	that will not that it is are yeah exactly
0:19:59	they were not
0:20:01	so
0:20:04	i
0:20:05	yeah yeah actually
0:20:06	therefore why have someone is that it is more interesting
0:20:10	to
0:20:11	yeah to thing to do that consideration
0:20:14	with in closing a right
0:20:17	because uh can one can include
0:20:19	yeah because we can define here at someone of quite
0:20:23	which is restricted on the position of the microphone
0:20:26	so only at a microphone position you
0:20:29	do you wouldn't
0:20:29	here anything
0:20:31	but in the listening position so for example if you can imagine
0:20:35	that
0:20:36	if
0:20:36	that is there a conference room and here's a table able with a microphone is on it
0:20:42	so it two would be with B
0:20:43	actually very practical
0:20:45	for the people to put their head on the table
0:20:48	the city
0:20:49	if they hear something on a
0:20:52	yeah that is actually i
0:20:55	a well thank you very much to little a is off
0:20:58	thus

SPATIO-TEMPORAL SIGNAL PREPROCESSING FOR MULTICHANNEL ACOUSTIC ECHO CANCELLATION

Echo Cancellation

Přednášející: Karim Helwani, Autoři: Karim Helwani, Sascha Spors, Herbert Buchner, Deutsche Telekom Laboratories / Technische Universität Berlin, Germany