0:00:13i build your presentation about rope and direction estimation method
0:00:18using component pressure and energy gradients
0:00:21and
0:00:22this work has been a with
0:00:23use that we a cue from
0:00:25although universe the also
0:00:29okay well here is to outline
0:00:31my presentation
0:00:33first uh short introduction to this topic
0:00:36and then some background
0:00:38about the direction estimation
0:00:41mean G chi i don't analysis
0:00:43and also uh
0:00:45i will present the microphone a right which probably used mike the
0:00:48that's signals for for this and now alice
0:00:52and then uh
0:00:53i will present the
0:00:54this big method
0:00:56for direction estimation
0:00:58this this come from the rich
0:01:00pressure and energy gradients
0:01:03and also the microphone error rate which is optimized for this method
0:01:08and uh then some evaluations and
0:01:11one of the summer of this presentation
0:01:15ah
0:01:15well
0:01:17T estimation of direction
0:01:19well it S of or or per pulse in several applications
0:01:23like a a source local station and beamforming
0:01:27and uh also in in uh
0:01:30this got of parametric spatial audio coding methods
0:01:34and that there's a huge
0:01:36or a large scale of
0:01:38you for that's this estimate direction
0:01:41like in music
0:01:42and it's breed
0:01:43it's cetera there are
0:01:44but here we are concentrating do
0:01:47the direct uh
0:01:48these sound in this the based
0:01:50may that's
0:01:52so
0:01:52we are using that's for
0:01:54for direction estimation
0:01:57and that
0:01:58this kind of approach
0:02:00has been used
0:02:02with the directional audio coding
0:02:04which is sir
0:02:06technique for recording and a repair routing spatial sound
0:02:11and a whole
0:02:13here in this figure you can see
0:02:15one the application
0:02:17teleconferencing
0:02:19where we have a
0:02:20some remote location
0:02:22there are some
0:02:23some twelve or send
0:02:24microphone array which kept to the sound and
0:02:27and then we do some
0:02:29some encoding and decoding
0:02:31and that
0:02:32then we should have a somehow
0:02:34spatialised telecon for from the other end
0:02:38"'kay"
0:02:42uh
0:02:43so um this uh
0:02:45noted noted and i'll analysis is based on the sound in those vectors so
0:02:50which uh it
0:02:52which are uh represent the direction and magnitude of the
0:02:55that's flow of sound energy
0:02:58and uh
0:02:59this uh vectors are
0:03:01are computed as a
0:03:03pressure at times particle well velocity in one point of sound field
0:03:08and uh
0:03:09oh the direction of the rival
0:03:11it's of obtain it
0:03:13a simply bleep taking an ops of a side
0:03:16opposite direction of the
0:03:17so sound to the vector
0:03:20and um you know or applications
0:03:23a related to do do you arc we have used to
0:03:26be format microphone signals
0:03:28in this analysis
0:03:30so this
0:03:31signals consist of
0:03:34of one omnidirectional signal on and three
0:03:37three die was four
0:03:38for X Y and chit
0:03:41directions
0:03:42so these type they
0:03:44the approximate the
0:03:46the body go well all C D's
0:03:50and uh
0:03:54uh
0:03:54instead of using a
0:03:56for instance sound field microphone
0:03:59or
0:03:59another
0:04:01and the microphones for for be form microphone signals we have
0:04:04you have been used this kind of
0:04:06uh
0:04:08microphone a rate of
0:04:09or four
0:04:10only direct sum microphones
0:04:13which are placed close to one another
0:04:15and up
0:04:17uh
0:04:18this a horizontal be format signals can be derived it
0:04:21from this this kind of error rate
0:04:24and uh the idols or
0:04:26computed just type biting you known
0:04:28taking a breast a gradient
0:04:30from opposing microphones
0:04:32so X type of the wide of are just
0:04:35you one want direct the would
0:04:37each two and
0:04:38and so on
0:04:40and uh
0:04:41well
0:04:42and this
0:04:43W signal this only direct something lights
0:04:45just a and number eight over
0:04:47or microphone signals here
0:04:51but the unfortunately we have some
0:04:53problems with this
0:04:55this kind of error rate
0:04:57when creating those those type
0:05:00goes at high frequencies
0:05:02uh this type was so deformed
0:05:04because of the spatial
0:05:06and that
0:05:08well this
0:05:10this uh
0:05:10the spatial realising frequency
0:05:13here
0:05:14if the depends on the and the distance between opposing microphones
0:05:19and here i have a a well that
0:05:22three different
0:05:23three
0:05:24three figures for three different erase
0:05:27so uh
0:05:30this
0:05:31oh
0:05:32first one here
0:05:34well this is for
0:05:36yeah or with the one centimetre be distance
0:05:39and uh
0:05:40well it it produce quite a it die balls
0:05:43that
0:05:44or or frequencies here
0:05:46but when we increase the distance between microphones and centre
0:05:51will be some problems at high frequencies here
0:05:54and here
0:05:55so these are not bibles anymore
0:05:59and up
0:06:01obviously
0:06:02this
0:06:03has some influence on
0:06:05on a direction estimation
0:06:08so at high frequencies and up
0:06:11uh here is the
0:06:12this a direction
0:06:14well the the estimation error here
0:06:16it's express it does uh root mean square error or here in this figure
0:06:22a a function of frequency
0:06:23so
0:06:25at high frequencies
0:06:26after this or specialising frequencies
0:06:29this uh
0:06:30yeah or is quite
0:06:32a huge
0:06:33and uh
0:06:35and on the other hand a low frequencies
0:06:38also depending on the distance between microphones we have some
0:06:41some estimation error because of the inter
0:06:44in a no of the microphones
0:06:46and uh
0:06:49and basically we can estimate the direction reliable
0:06:52only within a
0:06:53so then frequency window
0:06:59um um
0:07:02so um
0:07:03in this work
0:07:05we are proposing to use um
0:07:08this kind of a array
0:07:10which uh a consist of four
0:07:13four omnidirectional microphones with
0:07:16relatively large housing so
0:07:18so they are shallow so i won't have high frequencies
0:07:21and this
0:07:22this provides us some
0:07:24some in microphone level differences
0:07:27so
0:07:28uh this uh microphones are are run such that there one axis directions are pointing to the
0:07:34a post side directions
0:07:36in in this microphone pairs pairs here
0:07:40and uh
0:07:42well
0:07:43this perhaps just so just a sure rest of how
0:07:46how this uh
0:07:48as sound this shot what and uh at rated
0:07:51because of the chateau one so in this direction
0:07:54directional patterns here
0:07:56and these are for two two different microphones this left one is four
0:08:00eight K G microphone
0:08:02which is larger done this another one this grass microphone here
0:08:06but anyway we can see that
0:08:08these are not
0:08:09on directional direction anymore at high frequencies so
0:08:13and uh so
0:08:15uh
0:08:16this
0:08:17this effect
0:08:18this you the lies here
0:08:20here with
0:08:21direction estimation then
0:08:26and up
0:08:28so um
0:08:30um
0:08:31for for estimating direction
0:08:33we are
0:08:34proposing to use
0:08:36uh or or computing the energy gradients between those microphones are high frequency so
0:08:42it's up
0:08:44just computing the
0:08:46the
0:08:47subtraction between power spectrum of the microphones
0:08:51as that
0:08:52we are
0:08:53we are approximating sound in directly with this
0:08:55with this up action subtraction here
0:08:59and up
0:09:00it
0:09:01produce that's
0:09:02this kind of
0:09:03type will direct direct is for
0:09:07for a
0:09:08for this
0:09:10approximate a approximated in this to the vectors here
0:09:13and uh we are using directly D
0:09:16these for direction estimation at high frequencies
0:09:21and up
0:09:22well
0:09:23but on the other hand we don't have any
0:09:25any uh
0:09:27major or
0:09:28or we don't have a prominent
0:09:30inter michael level difference is that low frequency so there we use
0:09:34use just very shall make that for for computing first
0:09:37pressure gradient and then
0:09:39then in the C the vectors from them
0:09:42so this is somehow
0:09:44combination between impression that in you gradients
0:09:50uh
0:09:52okay well i
0:09:54uh
0:09:55then another
0:09:57i don't topic in his presentation was to you
0:10:00uh optimize
0:10:01microphone a rate for this
0:10:03it computation
0:10:05so the idea here is to
0:10:07knots
0:10:09this a spatial i freak ones with the
0:10:12frequency limit for using the energy gradient
0:10:16and uh
0:10:17so as i mentioned this
0:10:20this is a i frequency it's depends on the
0:10:24inter microphone distances
0:10:26and uh
0:10:28frequency lee for for in into gradients it's depends when the dive faq "'em" size of the microphone
0:10:34and uh
0:10:36here this um
0:10:38a a we no effect four omnidirectional microphone it's uh
0:10:42speech described with the
0:10:43directivity index
0:10:45which is a
0:10:47ray sure
0:10:48between uh on axis energy
0:10:51and a total
0:10:53so energy we just integrated over all directions
0:10:57of this
0:10:58you all some
0:11:00some idea about this direct sum no use of the
0:11:03omnidirectional microphone that high frequencies
0:11:06and uh on the other hand this uh a direct to be index it's
0:11:10depends on the ratio
0:11:12of uh between a die fry came circle for Ms
0:11:16and wavelength
0:11:18well this K A
0:11:19factor it's
0:11:20three that's this ratio
0:11:23and uh
0:11:25and uh after
0:11:26and would this
0:11:29and with this direct T V index and a K a factor we get this kind of
0:11:35this kind of gore
0:11:37for omnidirectional directional microphone
0:11:39so this this represent this uh directive the index
0:11:42as a function of
0:11:43K A
0:11:46and uh finally
0:11:47we can compute the
0:11:49optimist distance between microphones
0:11:53with this
0:11:54formal here
0:11:56so
0:11:57basically we just a
0:11:59defined that how much we want this up
0:12:02to use this data shadow a we affect here
0:12:05uh
0:12:07so we just
0:12:08choose one
0:12:09some
0:12:10that are direct to be index value here
0:12:13and then it we take the corresponding K of well oh here and then compute
0:12:17the distance
0:12:21okay well are
0:12:24then some evaluations
0:12:25uh this were
0:12:27conducted in
0:12:28and a and i a chamber
0:12:30on the measurements were done in and the chamber
0:12:33and that
0:12:34using a
0:12:35a K G microphones
0:12:37for a gauge you microphones with
0:12:39i for i come of
0:12:40two point one centimetres
0:12:43and that this results in a spacing of
0:12:46three point three centimetres for for this error rate
0:12:50and that
0:12:51also using grass microphone error right
0:12:54which has a more
0:12:56small die for kim size then
0:12:58this a K G microphone
0:13:01and uh again we have a
0:13:03this uh
0:13:04estimation error
0:13:06expressed
0:13:07as a
0:13:08root mean square or here
0:13:10so um
0:13:12this results this solid line
0:13:15this is for
0:13:17for using this uh the additional method using those rest gradient only
0:13:22and that
0:13:23well
0:13:24well as you can see that
0:13:26at high frequencies to zero or is quite
0:13:28it's very significant
0:13:31after just the spatial lies and frequency
0:13:34but
0:13:37but this energy gradient they produce the it's produce very nice nice estimation for us
0:13:43and uh
0:13:44and using a combination
0:13:46of these different
0:13:48radiance
0:13:49we get
0:13:50somehow
0:13:52uh
0:13:52reliable estimation for all
0:13:54for
0:13:55for entire
0:13:56what audio frequency range here
0:13:59and the same with this
0:14:00grass microphone array right
0:14:02you
0:14:06uh
0:14:08so um
0:14:09yeah the summary of my
0:14:11my presentation
0:14:13so uh
0:14:15so the basic idea was to
0:14:17to improve
0:14:18T and now this is which is
0:14:20the direction estimation
0:14:22from the from using this a square microphone error
0:14:26and uh
0:14:27see improvement
0:14:28has
0:14:29actually it by using a
0:14:32using the shot of the microphones and this make method
0:14:37and that
0:14:39and also it was shown that this
0:14:41optimized
0:14:42microphone from rate it's works with this spec method
0:14:48okay well
0:14:49thank you
0:14:50i Q and have time
0:14:56the question
0:15:04i i i think about
0:15:05a work the way
0:15:07which
0:15:09and
0:15:10right
0:15:12i
0:15:12were
0:15:13have you ever cut the experiments i mean pressed of reverberation
0:15:17you should the experiments results any quick humour
0:15:21no
0:15:22but
0:15:25oh right
0:15:26do you have any uh yeah experiments
0:15:29uh a and the experiments in in way easy or or really environments
0:15:33oh yeah yeah yeah i have i have a i have a
0:15:36yeah tried this with this a a teleconferencing application
0:15:40and uh
0:15:41you
0:15:42well
0:15:43it works
0:15:43nice
0:15:45this in
0:15:46well i i i'm this in our experiment of that have we have used it in or more room and
0:15:50also some
0:15:52much environments
0:15:53so yeah
0:15:57i and more questions
0:16:02okay thank you very
0:16:03you