Přepis řeči - CONVERTING 5.1 AUDIO RECORDINGS TO B-FORMAT FOR DIRECTIONAL AUDIO CODING REPRODUCTION

0:00:13	thank
0:00:14	so i make a bit like and uh
0:00:16	and i will present our paper called converting a one one all audio recordings to be a format for directional
0:00:22	audio coding with blocks
0:00:25	right long a let's score
0:00:27	so yeah yes line of the presentation so i was that
0:00:31	like to be an introduction to the topic
0:00:34	and then give you some background information about this and then the actual something
0:00:39	and i will and with some conclude
0:00:43	so yeah
0:00:44	several parametric spatial audio coding methods
0:00:47	for example and pixel round or a parametric stereo
0:00:51	it's that a
0:00:52	but
0:00:53	and they are really good but to become a they have some certain kind of reproduction system for some stereo
0:00:57	or five point one
0:00:59	so it would be perhaps nice if we could have some kind of
0:01:03	oh any kind of in to the
0:01:05	system and and it would also reduce the sound using any
0:01:09	reproduction system but some any loudspeaker layout or what sample head
0:01:16	so what you have in deal you've ever been in our universe these this thing all the directional audio coding
0:01:21	all the know
0:01:23	and
0:01:24	this this thing to use this as the form in input
0:01:27	and you can use any loss speaker layout as a
0:01:30	for reproduction and also had poles
0:01:32	so we think that the uh that would be perhaps a suitable can that for this kind of generate
0:01:38	only a form
0:01:40	and in this study
0:01:41	i will
0:01:42	think about
0:01:43	what that for example white one one or something that is really common on no more well
0:01:47	multichannel reproduction so i ones
0:01:49	uh
0:01:50	study about and of buy been studying about
0:01:53	could we can will convert these five one one signal that's to be for
0:01:57	and we would be we them with a directional audio coding without that any artifacts
0:02:03	and then
0:02:03	in this case your case for example
0:02:06	we tried to
0:02:07	not a point one one and get a set and that of reproduction of a sets of sound
0:02:12	and get the same thing
0:02:13	through this can of and and the uh these directions audio coding reproduction production
0:02:20	so
0:02:21	first i will so you something about this
0:02:23	the are technology so the basic idea and is it
0:02:27	that we don't try to reproduce the sound really physically for right
0:02:31	instead we tried to reproduce it in a way that it would be perceptually correct so how we perceive does
0:02:37	some real would be the say
0:02:38	as in the original space
0:02:40	and that
0:02:41	possible see its for this technology would be for simple telephone thing
0:02:46	and also spatial sound reproduction
0:02:50	and that's i set
0:02:51	the input to this
0:02:53	uh
0:02:53	processing is the form
0:02:55	three
0:02:56	so it contains a the directional signal W
0:03:00	and and die X Y and set
0:03:02	which uh
0:03:03	direct that along the
0:03:05	coding case of a to connotation that's just
0:03:10	so what we
0:03:12	so to is that we process the sum some in time-frequency frequency domain
0:03:16	and why we do this is that we tried to me the temporal and also the frequency resolution of of
0:03:22	yeah
0:03:23	so we try to do with the same resolution solution that's our
0:03:27	hearing system works
0:03:30	and
0:03:31	next let's think about what should we and the last of the sound
0:03:35	so we are do the analysis for each time for it's time to get is
0:03:39	things without L
0:03:41	uh
0:03:42	much to the solution
0:03:44	and uh
0:03:45	first would be to do is we analyzed the direction of sound so so simply just wait it coming for
0:03:49	left right up don't
0:03:51	so on
0:03:52	and then we have this another parameter was called diffusion as
0:03:57	which is this uh
0:03:59	means that
0:04:00	is the sound coming only for certain direction is it like a minute weight
0:04:04	or always it got that's completely diffuse coming from all directions but like for example reparation to be
0:04:09	or is it something in a we
0:04:14	and yeah is the scene that is part of processing and the whole block diagram of we you can see
0:04:20	so
0:04:22	as you get see yeah
0:04:23	we done by this time we is the composition for simple using a short time fourier transform
0:04:29	and then we do the processing separately for each frequency that
0:04:33	so we get use a direction analysis
0:04:36	results from yeah
0:04:38	and what we do that the old your that is that we create these kind of work so might of
0:04:41	course which can be a sample
0:04:44	uh card us to was lots to a store this really similar to first order ambisonics these
0:04:50	these H
0:04:51	but the take what we do is that we divide the sum of two different
0:04:55	it's is for known used three
0:04:57	and the is used to be
0:05:00	and we break reduce them you know different way
0:05:02	and for the don't need used three we assume that it
0:05:05	basic a for of the direct sum so we try to
0:05:08	rep reduce it that's a point like source using for example i'm to to binding or in three D case
0:05:13	vector base
0:05:14	base and it would kind
0:05:16	and
0:05:17	for the diffuse if you stream we ask them that it's rounding so for example real so what we want
0:05:22	to get a get a kind of
0:05:24	and by her section so what we want to use a for some but the correlation to get it
0:05:29	running
0:05:30	and in the and we use these two streams and also the all frequency bands and we get a a
0:05:34	a a uh
0:05:35	lots a C
0:05:39	and
0:05:40	using this made up with the in it that's been studied for a couple of years and you can get
0:05:44	quite nice audio for for real recordings with a a form a microphone
0:05:50	so let's scroll next to the actual topic what by we get a like what we were studying in this
0:05:56	right
0:05:57	so how to really convert this kind of wine
0:06:00	five point one audio recordings to be much so at first it seems really seem also
0:06:06	what we simply can do which we problem from this kind of well lot record
0:06:10	so we position the loudspeakers according to itu standard and then we just assume that we are in a a
0:06:17	it works space
0:06:18	and we perform kind of words the recording with the people microphone
0:06:22	is really simple be can use simple cosine on a sine functions for this
0:06:28	but
0:06:29	however there were some problems that we power
0:06:32	so if if about how
0:06:34	human
0:06:35	yes this kind of reproduction if you have some
0:06:37	the if use of a reverberant something part one one case
0:06:41	we perceive is it as the diffuse and around
0:06:44	but if you think about how
0:06:46	there there a nice is
0:06:47	analysis
0:06:48	works in this case
0:06:50	as i said before that
0:06:52	the diffuse nest is one you we have
0:06:54	sound coming E from all directions and also these is
0:06:58	i think the
0:07:00	is also true in that you think about what is stiff use field
0:07:03	but in this case we have a most of the last because i in the problem
0:07:07	so we don't have equal energy coming from all directions and i so
0:07:11	a result we have a low weight use nist and one
0:07:15	and what this course is that the we get the by guess of sound source the to send the lot
0:07:20	be
0:07:21	and it's not received to be completely diffuse
0:07:24	and surrounding the solve
0:07:26	that this is something of course that we don't cool
0:07:30	that we got the second idea how to fix this problem is that we isn't the speakers equally around the
0:07:37	we form a microphone and
0:07:39	this is
0:07:39	was
0:07:40	working quite nice that the now you get the diffuse sound coming from all directions T
0:07:46	and you get to correct used as well
0:07:50	and also you that quite nice all your
0:07:52	quality but the problem is of course that
0:07:55	no the loudspeakers a
0:07:57	it wrong basis because if you think about do you should have a
0:08:01	sound so that what simple minus the decrease but now we have a feature that sound source here
0:08:06	so we get kind of a by in the reproduction
0:08:09	which we of course can
0:08:11	point to by in the reproduction side that assume that we have a speakers at those
0:08:15	direction
0:08:17	but we might have also really people my reporting
0:08:20	which you would might perhaps too much much these uh one but it seeing as and in this case you
0:08:25	would have a wrong directions in i don't of
0:08:28	case
0:08:29	so we need something more
0:08:32	a lab right to get is
0:08:33	as function in all that the time
0:08:36	and then we have a
0:08:37	i will present this
0:08:38	but that that we propose a would be that a
0:08:41	so what we do is that for the directional provost
0:08:44	some for example
0:08:46	a sources or direct sound of the
0:08:49	uh
0:08:50	sound
0:08:51	we use the standard layout so for that part we get the directions right and that is
0:08:56	of course the important in that
0:08:58	and if we have some reference some then we used even layout
0:09:02	to get the diffuse is that you correct
0:09:05	and this is really
0:09:06	nice in a way because the both directions and the diffuse nist a correct when they are the most in
0:09:11	four
0:09:13	oh let's look at the block diaphragm
0:09:15	so
0:09:16	we do the analysis
0:09:18	and
0:09:19	a a frequency domain so we start by these time because of the composition
0:09:24	and what we do is we can you these
0:09:27	we watch other be not signals use the even and layout and the standard let loudspeaker layer
0:09:35	and what we do is we simply a these two three is a holding to the D C use ms
0:09:40	of the summer
0:09:41	of the some you and we do that
0:09:43	analysis using these you and lots of light
0:09:46	and we can see
0:09:48	cross fade or was played between these two three
0:09:52	to get the most
0:09:53	the
0:09:55	people three
0:09:57	and we have seen that these would be quite close to
0:09:59	but it should be ten
0:10:00	next we what we do the and less the diffuse as again
0:10:04	if this is correct now
0:10:07	and we
0:10:07	see from here that
0:10:10	if we think about this
0:10:12	re line each the diffuse this can do with that
0:10:15	even a speaker layout we get
0:10:17	correct it used
0:10:18	where with this standard layout we would get
0:10:21	to load diffuse noise
0:10:22	so we can see that if we much
0:10:24	we we these two
0:10:26	streams
0:10:26	we get
0:10:28	correct if used as if it's really high or really low but in the be no be get a kind
0:10:32	of a little bit and the estimation
0:10:35	so we need to by a big for
0:10:39	so let's look got this if use this take of once small the see
0:10:42	would be a would be more five to get a correct if used
0:10:46	so we can see yeah that this is the uh
0:10:48	pressure and this is the part of as
0:10:51	so we can see that these if these are quite P well
0:10:55	the whole term yeah is
0:10:56	white white and you get that all that the use this is really low
0:11:00	and so
0:11:01	if that
0:11:02	a particle velocity
0:11:04	well basically the dipole signals and all me that's of uh he but we get a low here
0:11:09	and on the very if you have if these two
0:11:12	so it's a
0:11:14	i
0:11:14	different we can
0:11:16	really low uh
0:11:17	but for this um so we get right
0:11:21	so we can simply money by the diffusion is by
0:11:24	at just in the ratio between the only a directional signal and the dipole signals of the beep
0:11:29	oh this is actually done
0:11:31	you need a bunch of equations and i would not give you hear but you can read from the paper
0:11:35	if you in
0:11:38	so let's go back to the this
0:11:39	block diagram so we can see that we on but the that if used as well use and we get
0:11:44	this kind of a
0:11:45	uh coefficients for access to the
0:11:47	these used
0:11:49	and it the next stage we simply
0:11:51	molded by uh the the so signal and the died
0:11:55	signals so we get the final
0:11:58	we almost three
0:12:01	so we can use them to a use these thing as reproduce the sound we get a and we also
0:12:06	can it with the other the real people are recordings for example
0:12:11	yeah we didn't do
0:12:12	for only saying this yet but at least according to in formal test we didn't
0:12:17	we didn't find to be put course any degradation of of audio or you probably
0:12:25	so
0:12:25	as of convolution
0:12:27	that audio an audio coding or get out
0:12:29	is a perceptually motivated method to reproduce spatial sound
0:12:34	and it has been a before using mainly be you for my or us that in
0:12:39	and so in this paper or the idea was that could we use also for simple life one S to
0:12:44	in
0:12:45	to these
0:12:46	uh reproduction method
0:12:48	in order to have for example these kind of check audio for
0:12:51	and we
0:12:53	showed to use i thought use a two simple matrix in based methods
0:12:57	but they had some problems both of
0:13:00	and the proposed method
0:13:02	in there we analyse the virtual sound field generated by the five point one or
0:13:07	and we do the analysis in the time frequency my
0:13:11	uh
0:13:13	and by using this we were able to do these kind of people
0:13:17	a and that we did not find two
0:13:19	uh the rate the audio quality
0:13:22	and also we
0:13:23	and uh
0:13:24	makes the resulting be not that's with other be
0:13:28	signals
0:13:30	yeah some acknowledgements and
0:13:32	i would like to check
0:13:41	yeah
0:13:46	i
0:13:55	i
0:14:00	hmmm
0:14:00	i
0:14:01	i
0:14:05	i
0:14:06	more
0:14:08	yeah
0:14:08	yeah
0:14:09	i
0:14:11	this
0:14:12	in physical they are not a at all the say what perceptual at the same at these like one to
0:14:17	lower in format test so if you look at B Y so it's a completely different but it some the
0:14:21	say and that's the uh a little that it puts some the same
0:14:26	uh_huh hmmm
0:14:28	oh
0:14:30	well what what we wanna do this is that what simple we would also read produce using gonna use in
0:14:34	must every all seven what one O any
0:14:36	a lot of layout are also had also so that's why we want to get also the be or more
0:14:41	uh this five one of these people bust
0:14:48	right
0:14:49	i
0:14:51	oh
0:14:53	oh
0:14:54	no that
0:14:55	directional audio coding and
0:14:57	also
0:14:58	and also the reverse and labeled
0:15:00	on a a bit twice
0:15:02	same so is it's only for sets choice say
0:15:10	i
0:15:13	well
0:15:14	oh
0:15:14	i
0:15:16	oh
0:15:17	is
0:15:19	yeah i
0:15:20	one
0:15:21	i
0:15:22	wow
0:15:23	oh
0:15:24	i
0:15:24	you
0:15:25	uh sweet spot of sweet spot
0:15:27	is used for
0:15:29	no i i
0:15:30	it's of a goalie this it does make a big difference at least we didn't for a of any brought
0:15:35	class but of course if you go away from the of three spot you are more easily would you to
0:15:39	get a a a plus but we mainly due the since text
0:15:42	this
0:15:42	the sweet
0:15:44	but there was really noise
0:15:51	uh_huh uh_huh uh_huh
0:15:55	hi
0:15:56	i
0:15:57	okay
0:15:59	and
0:16:00	this
0:16:01	oh
0:16:02	i
0:16:04	i
0:16:04	hmmm
0:16:06	no uh we really did i don't
0:16:09	see
0:16:10	of course we could
0:16:11	because we don't wanna at any room at that basically because we want go present but
0:16:16	yeah of course
0:16:18	yeah
0:16:19	i don't know how do we really
0:16:20	get it in it but yeah
0:16:27	i
0:16:37	uh_huh
0:16:40	i did it in
0:16:41	room which is a a point to this decide to use that that so it's list
0:16:45	what there
0:16:46	white nice the i we need some really brief test also in nine a will and that we didn't find
0:16:51	a but was also that but
0:16:53	the main main part of the test was done in really we
0:17:00	okay
0:17:02	know
0:17:03	i
0:17:04	yeah

CONVERTING 5.1 AUDIO RECORDINGS TO B-FORMAT FOR DIRECTIONAL AUDIO CODING REPRODUCTION

Spatial and Multichannel Signal Processing

Přednášející: Mikko-Ville Laitinen, Autoři: Mikko-Ville Laitinen, Ville Pulkki, Aalto University, Finland