Speech Transcript - EFFICIENT RANGE EXTRAPOLATION OF HEAD-RELATED IMPULSE RESPONSES BY WAVE FIELD SYNTHESIS TECHNIQUES

0:00:13	actually a and introduction
0:00:15	and the will um
0:00:17	so
0:00:18	some part of the motivation was already given by D
0:00:21	so um head related transfer functions or impulse response
0:00:25	characterise the was
0:00:27	a from source
0:00:29	measure
0:00:30	for example from a loudspeaker
0:00:31	the years of a i mean that really
0:00:34	or us and uh is your microphones
0:00:36	and
0:00:37	these functions are known that they are of course
0:00:40	students angle uh dependence was you for example would take you had uh i if a and they also to
0:00:46	some extent at distance
0:00:48	so
0:00:49	for example to use to situation gets free does me to does of course a big difference in and uh
0:00:55	H T S
0:00:56	if you know want the set up something like for example of a well what will auditory environment
0:01:01	then uh
0:01:02	it would be nice to have a database of hrtf covering all potential positions but like to have a which
0:01:08	was also
0:01:09	and of course the measurement of such a database is a quite tedious files
0:01:14	so people can up with the idea of
0:01:16	if you know
0:01:17	for example of the hrtf a a set for all incidence angles for one distance
0:01:22	how can you
0:01:24	compute or
0:01:25	syntactically we create a yes different
0:01:28	this is
0:01:29	and that
0:01:30	these techniques are on a range extra for
0:01:33	and of course that some techniques which uh
0:01:35	uh uh based on psychoacoustic and some based and physical
0:01:39	assumptions and i will uh are physically degraded
0:01:44	of course and the past the a lot of uh stuff has been done here uh for example uh
0:01:49	uh
0:01:50	at for it to such a can measure a must you also circle
0:01:54	so it's quite natural to sing yeah that the could expand them into spherical harmonics
0:01:58	it's not that straightforward as one would assume you need to exploit you the prosody uh of a big the
0:02:04	creation or to do so
0:02:05	and then you can expand on the uh data set
0:02:08	terms of so if a spherical harmonics and
0:02:10	spherical hankel function of of the form
0:02:13	uh range
0:02:15	one problem image of a person in the context of very from one is that if you have a uh
0:02:20	a not very dense sent think for it or even if bits on forty but like this is coded weekly
0:02:24	data sets
0:02:25	then you these methods are quite prone to numerical instability
0:02:29	okay a stuff and that's of course some
0:02:32	uh
0:02:33	"'cause"
0:02:34	but a class of texas based on the uh
0:02:37	idea of uh what will acoustics and i will go to detail a time
0:02:41	and the idea here is that you can X
0:02:43	uh interpret
0:02:44	yeah data set as of what will also be carry
0:02:46	of course if you would take all speaker
0:02:48	had a a a and it room that's the same
0:02:51	have
0:02:52	a loud speakers
0:02:55	and of the come back to that
0:02:56	so uh if you have a for example and hrtf a set which just examples here
0:03:01	quite costly
0:03:02	for the straight you have a right i H the F and we have all of these
0:03:06	and
0:03:07	yeah
0:03:08	so if you can that this as an ensemble of lots because what will of course be oh
0:03:14	uh the data
0:03:15	then this is an speaker
0:03:17	if you know for for instance
0:03:19	want to have a H T F's from and
0:03:21	the difference source of the distance
0:03:23	like like from yeah
0:03:24	uh what you can do is uh
0:03:27	you can think
0:03:28	but this we can interpret this as and what will uh
0:03:31	so some what will some system
0:03:33	so you have a a T yes three percent a lot i rate
0:03:36	drive this
0:03:37	what's will loudspeaker read
0:03:38	by an appropriate a in signal
0:03:41	which as
0:03:42	oh
0:03:44	so at first sight one saying
0:03:45	oh
0:03:46	uh uh and some um this is the head of the that's not
0:03:49	a but or by the take
0:03:51	and that's
0:03:51	only partly true uh uh and this context
0:03:54	play role
0:03:55	course
0:03:56	if you for example have
0:03:57	scattering the head but the source but real source
0:04:00	or
0:04:01	uh
0:04:02	synthesized sound correctly
0:04:05	yeah doesn't care or something comes
0:04:08	it's easy for them
0:04:11	of course the sample synthesis introduces some errors
0:04:14	you
0:04:16	and the overall uh syntactic hrtf is then given from as the transfer function from the work clothes all over
0:04:22	the end
0:04:23	right
0:04:24	this
0:04:24	that
0:04:25	so in a minute how was looks in mathematics
0:04:28	so uh that's the two dimensional case here so uh assuming that we have taking measurements
0:04:33	on a circle around to is not that
0:04:37	that's not a free field case so don't hrtf
0:04:40	and it's on of is this you typically start
0:04:42	with a continuous
0:04:43	diffusion or
0:04:45	so called secondary so
0:04:46	but kind of this continuous loudspeaker
0:04:49	that you wait
0:04:50	right
0:04:51	few into all over all
0:04:53	you use a a the end up
0:04:55	pressure
0:04:56	at all points
0:04:57	yeah
0:04:58	and uh to be is assume that you have a point source located of point source someone a whole model
0:05:04	for you second a resource that's given by the
0:05:07	so if we now want to load the uh
0:05:10	do you pressure at the left and right yeah
0:05:12	it's straightforward forward to replace the queens funk
0:05:14	a the uh a two data transfer function
0:05:17	the cost data transfer function is
0:05:19	if a measure it was a lot people also to
0:05:21	from from a point source
0:05:23	to one of the S
0:05:25	so it's quite straightforward to
0:05:27	like
0:05:28	and the overall transfer function is them from the source U S as a total for a you have to
0:05:33	take yeah just case that to driving function
0:05:35	uh should of course
0:05:37	a and what was also has to be driven by a to work well
0:05:41	but the region
0:05:42	transfer
0:05:44	um so for i can that the continuous phase of course you have to sample saying so you have a
0:05:49	a sent here
0:05:51	and uh in this case of course
0:05:53	it's not
0:05:54	extract any on you have
0:05:55	some kind of approximation
0:05:57	and on your space sending it the well
0:05:59	you will need and just
0:06:00	say me for send data sets
0:06:03	for what
0:06:04	then driving function
0:06:06	and let them right
0:06:07	H
0:06:08	and of course like always and sampling you may get uh
0:06:13	spatial at yes and and of a come back
0:06:16	so a of no i didn't talk about the technique used to calculate a drive functions and the construct a
0:06:22	takes around
0:06:23	uh which we term a sound field synthesis techniques takes that of techniques
0:06:26	uh
0:06:27	at at least it this is of a sound field
0:06:29	was was not extend the every using an ensemble law
0:06:32	so
0:06:33	physical synthesis this is that as the
0:06:35	basic assumption of sort that synthesis
0:06:37	and well known approach here of for example a few synthesis
0:06:41	you you compensated by the sonics
0:06:43	the spectral do method and a number of them are debate
0:06:47	for
0:06:48	and but you can use use all of these approaches here
0:06:51	not to
0:06:52	voice
0:06:53	for more less
0:06:54	i we can of a few synthesis for some reason that see
0:06:57	later
0:06:59	so a very brief introduction to but if it's and this is uh
0:07:02	if you want to know more about because in this is i would like to do a you to the
0:07:05	paper is to time
0:07:07	train
0:07:08	a have a make some this is is based on an approximation of the synthesis
0:07:12	so if you have yeah
0:07:13	a sound field of the but will solve
0:07:15	can can approximate this by
0:07:17	this integral over a
0:07:19	oh is so face
0:07:21	yeah we are you want to synthesized
0:07:23	a sound field and
0:07:24	you have a
0:07:24	as at
0:07:26	which is a window function you have a directional gradient
0:07:29	a some of you
0:07:31	false that's the
0:07:32	gradient direction of all the vector
0:07:34	and in
0:07:35	for
0:07:36	so this part is right from
0:07:38	a function a
0:07:39	takes care that only part of the because i used
0:07:42	and synthesis
0:07:43	and typically that hot
0:07:45	a
0:07:45	the but at all
0:07:46	propagate into the action of the
0:07:48	but was
0:07:50	um some properties of a it's of this is uh just summarise here that are some minor deviation
0:07:55	you to be in both approximations a
0:07:58	estimation yeah that's some
0:07:59	for few assumption
0:08:01	but it so for a very low frequencies and nearby by sources it's
0:08:04	has some
0:08:05	minor deviation
0:08:07	if is this is typically a model based approach which means that you have some models for you but was
0:08:11	also
0:08:11	but good in our context yeah
0:08:13	so plain based on sources
0:08:15	i have some lakes or you and uh just
0:08:18	calls
0:08:19	and
0:08:19	as we a back and i'm minute later or you can have source inside to the thing every so called
0:08:24	and a very important fact
0:08:26	for this paper is that
0:08:27	and a if it's in this is allows for very efficient implementation using a pretty for the ring of the
0:08:32	but was source signal and in the for the out
0:08:35	so you and really
0:08:36	meant to the time T the time domain be efficiently
0:08:39	and is in a and in as they no regularization
0:08:44	so um some pictures just the depict would make it and is is going in this cost you i simple
0:08:49	it a political what will uh system
0:08:52	this is uh uh close and hrtf dataset this
0:08:55	five five sampling
0:08:56	that one meter
0:08:58	so it can see for but will one source
0:09:00	but at three meters
0:09:01	for one thing about
0:09:02	see C you just synthesized
0:09:04	correctly so as to
0:09:06	oh that it is
0:09:07	and a four is we can see
0:09:10	C yeah that's the
0:09:11	spatial aliasing a so some
0:09:13	for things happen
0:09:14	you
0:09:15	spain
0:09:17	um that's being some so some on the uh certain properties of spatial aliasing and uh from him a basic
0:09:23	and for typical of a synthesis set
0:09:25	might be perceived will be to spatial aliasing in
0:09:29	the
0:09:30	however
0:09:30	if a synthesis all some systems also not sense of was that high
0:09:35	number
0:09:37	oh back to focus or since that even want first thing than the context of uh
0:09:41	spatial something
0:09:42	and a focused source is of
0:09:44	uh
0:09:45	that's the kind of a if you can generate a such techniques uh i
0:09:49	if if it's and also
0:09:50	but you
0:09:51	for was you if you to two a focus and so if a if you converges
0:09:55	the
0:09:56	just point and directors off
0:09:58	and that is focus point
0:10:00	we have the few
0:10:01	a more as a point so
0:10:04	and and uh an interesting property is that if you increase the frequency still give it to cut outs have
0:10:09	four khz
0:10:10	you can see that uh
0:10:12	the area which is a it is an scenario to too small small it is in free around the sensor
0:10:17	or point focus source
0:10:19	and he can four eight khz
0:10:21	we are
0:10:23	a i mean to be a inside
0:10:25	this is a
0:10:26	but a uh
0:10:27	so if you that
0:10:29	feel was all spatial
0:10:31	um and lot a property which is important uh i set already didn't do this in every has limited so
0:10:36	need if you
0:10:37	the low the focus point has
0:10:39	probably propagation direction or whatever
0:10:41	change
0:10:41	propagation gaze direction by changing but all speakers to use
0:10:44	in this example of it is also
0:10:46	side
0:10:47	oh
0:10:50	so uh you to but to use uh in terms of spatial that is the and in the following of
0:10:54	show
0:10:55	results which are related to focus source
0:10:58	this has to be used on the one hand at is and is not particular about
0:11:02	and on the other hand
0:11:03	this is this this and if you H T F to so interesting
0:11:07	because uh
0:11:08	H H G T S a change from dramatically for sources which are close to that this
0:11:12	so for two or three meters the a story
0:11:15	a we by what
0:11:17	to go for example half a meter or you you trying to see as the changes that what brought so
0:11:21	that's quite interesting
0:11:22	eighteen
0:11:23	since if you
0:11:26	so
0:11:27	yeah a some results uh the how to the it that that a straightforward uh uh use the time domain
0:11:32	implementation of but it's and this is so just some waiting and a with the so that really straightforward
0:11:38	a little more tricks a a and that things about both
0:11:41	oh have a this this
0:11:42	create realisation of the bit i'm going to
0:11:45	a more detail no
0:11:46	what if a if you want to and thing you have that that
0:11:49	yeah is that you're spatial exactly
0:11:52	and also
0:11:52	the distance
0:11:53	but also
0:11:54	front of you to some button
0:11:57	and and S um
0:11:58	a systematic and to if it is to come back in a minute
0:12:01	and uh for evaluation and used a love we uh a data set of H
0:12:05	um range uh
0:12:08	you can also
0:12:11	and it shows some results from you
0:12:14	so this is you one this is that in is not uh from
0:12:17	and source
0:12:19	denoted
0:12:20	i
0:12:22	yeah or note
0:12:23	to uh uh uh uh this uh
0:12:26	to use "'em"
0:12:27	S
0:12:27	and
0:12:28	if you consider that case is that you have a a S and homes on the plane
0:12:32	uh this is a a a a so for the how the man
0:12:36	oh the you the issues would assume
0:12:38	that uh
0:12:39	the also because would you look at a of you in if we dimensional space that's not the case for
0:12:44	or something plane and you know i'm about
0:12:46	so that lot that around you
0:12:48	and the also some
0:12:49	physically error
0:12:51	and
0:12:52	physically motivated has nothing to do with that
0:12:55	so um
0:12:57	and you have some systematic and deviation
0:13:00	or shown but no yeah for example if you have the you have a plane wave travelling
0:13:04	from the from uh from up to down
0:13:07	you can see a course a and B the polls years three you this well like
0:13:12	to have it
0:13:13	oh we have a if the and example
0:13:15	a
0:13:16	this site here
0:13:17	he that is some and
0:13:18	to to uh
0:13:19	uh
0:13:20	"'kay" should not assume of the head
0:13:23	so uh
0:13:24	yeah compensate for
0:13:26	that's score
0:13:26	important
0:13:27	if you would not do so
0:13:28	and level difference i i feel a uh wrong and i D is are very well for you uh for
0:13:35	perception of you
0:13:37	here
0:13:38	co four
0:13:40	so some results
0:13:41	um
0:13:42	that's is the one of the age of them
0:13:45	related impulse responses that uh
0:13:47	for for a four
0:13:48	each other
0:13:49	a
0:13:50	using uh and a data
0:13:53	P
0:13:53	a one
0:13:54	and
0:13:55	and C
0:13:56	quite nice
0:13:57	a a a a a a a major data set of how how meter
0:14:01	yeah
0:14:01	and see some
0:14:02	the
0:14:04	or here
0:14:05	there are a mind a and if you listen to the data set
0:14:08	yeah most
0:14:09	so the question
0:14:10	for
0:14:11	uh
0:14:13	that's
0:14:14	well for and a direct used the control level difference
0:14:18	um
0:14:18	results it is well known that the i'll likely be used for you by also so for example
0:14:24	yeah
0:14:24	right
0:14:25	the line and
0:14:26	yeah you do line
0:14:28	shows to measure data sets for for a good you and for how you can see that i i E
0:14:34	increases what we want and uh using a whole
0:14:37	except at at and also see
0:14:39	that uh uh i i be also increases
0:14:42	almost a low
0:14:43	to we measured one of a me to just did he is missus
0:14:47	might you to the fact that the or measurements
0:14:49	that all speaker
0:14:50	for half the measurements
0:14:51	some you the fact center
0:14:53	yeah
0:14:55	well
0:14:57	so it's also known that for nearby source is the low frequency response of the system
0:15:03	from increase in low frequencies from source is
0:15:06	that's shown here so that again
0:15:07	the dashed line
0:15:09	the the measurement a red fine
0:15:10	i
0:15:11	X one point at three the how i mean
0:15:14	you one
0:15:14	the
0:15:15	a made of one mean you can see
0:15:17	we can model
0:15:18	yeah use low frequency increases the
0:15:21	and also the overall all frequency one uh
0:15:24	a the little
0:15:25	of course there's some issue
0:15:28	sure
0:15:29	i
0:15:32	so some in
0:15:33	oh
0:15:34	i see that the application of
0:15:35	so so but will wave field synthesis
0:15:37	a matrix relation
0:15:39	but especially especially are correct results
0:15:41	the the form you know you're
0:15:44	and and uh
0:15:45	for for for of the for all the distances is that the false also works
0:15:49	get some spatial aliasing artifacts a question as for a C there are
0:15:53	and that's not
0:15:55	it
0:15:55	in number and the number they and the small realisation no involvement must in in the
0:16:01	but we might in that case it that's very low
0:16:03	and
0:16:03	big city
0:16:04	and you can also use
0:16:06	although the what was models and point sources
0:16:08	right
0:16:09	so i this is if you sources
0:16:12	and if you like to hear how it sounds than a mode some this in a lot more
0:16:17	form inside i'm sources
0:16:19	yeah are very minor differences between them and
0:16:22	hi
0:16:24	um i'm that some way to increase the uh
0:16:27	speech at is we can see
0:16:29	it can use multiple make it this is a present that this is the last week and the S
0:16:34	oh
0:16:35	a four
0:16:35	and of course uh would be very interest
0:16:38	taking a uh
0:16:39	if you to fixed using to models
0:16:41	have
0:16:42	i some this
0:16:43	i
0:16:44	a
0:16:45	and
0:16:45	for
0:16:46	oh actually a
0:16:48	i
0:16:49	i
0:16:50	i
0:16:54	i
0:16:55	oh
0:16:56	oh
0:16:57	one
0:16:57	oh
0:16:58	and that
0:17:00	my
0:17:09	yeah
0:17:11	yeah
0:17:12	oh
0:17:15	i
0:17:18	yeah
0:17:19	i
0:17:19	i
0:17:24	well
0:17:25	oh well
0:17:27	i
0:17:29	i
0:17:31	no
0:17:32	yeah and is not not to ring
0:17:34	for calculating you ones so if you have a data
0:17:37	a measure one and we meeting
0:17:39	uh
0:17:40	you can have for example assume that the data set for however
0:17:43	that's
0:17:43	quite a different
0:17:44	you can really here
0:17:45	do this
0:17:46	uh
0:17:47	and
0:17:48	a a and then the think examples can really here that there's a big difference with
0:17:51	we we don't have a
0:17:52	really here
0:17:53	a coming close
0:17:55	so
0:17:56	and otherwise it would have to matter all this this this which is uh
0:18:00	take
0:18:00	a
0:18:01	oh
0:18:03	that's the main make sure i play a
0:18:06	they
0:18:07	well
0:18:11	oh
0:18:15	no
0:18:16	yeah
0:18:20	or
0:18:24	i
0:18:26	house
0:18:28	yeah
0:18:29	yeah yeah and yeah
0:18:31	hmmm so
0:18:32	right
0:18:33	where
0:18:34	where
0:18:35	yeah
0:18:38	uh_huh
0:18:43	yeah that that's also the reason why we
0:18:45	and
0:18:46	because of you know that very must yeah that's small was the way to do it
0:18:50	or or or or also here are already that a spherical harmonics and uh and that's
0:18:54	how
0:18:55	and the you it's a more
0:18:58	correct
0:19:00	yeah okay
0:19:04	okay

EFFICIENT RANGE EXTRAPOLATION OF HEAD-RELATED IMPULSE RESPONSES BY WAVE FIELD SYNTHESIS TECHNIQUES

Spatial and Multichannel Signal Processing

Presented by: Sascha Spors, Author(s): Sascha Spors, Jens Ahrens, Deutsche Telekom Laboratories, Germany