Přepis řeči - Speaker clustering via the mean shift algorithm

0:00:06	uh so hi everyone um the most common
0:00:10	and that's and thank god university of uh
0:00:12	where is
0:00:13	and i'd like to talk about in your problem
0:00:14	to speaker clustering
0:00:16	uh
0:00:17	which
0:00:17	based mainly on some counts
0:00:19	so
0:00:20	um
0:00:21	say uh
0:00:23	individual ornaments analysis
0:00:25	uh so
0:00:26	that's when trying to be with these uh transition somehow
0:00:30	so um the outline
0:00:34	i
0:00:38	so
0:00:44	we should make
0:00:45	good
0:00:47	so uh i'm gonna talk about a little bit of but no parameter density estimate
0:00:50	and then i'm going to supply based on this the baseline and see
0:00:54	and then across a wide requires adaptation
0:00:57	uh in order to be compatible with a problem
0:01:00	uh and i'm gonna so some bass and setting would
0:01:03	i'm gonna show it like like that
0:01:04	is nothing more than say a to see
0:01:06	the modes of the posterior
0:01:08	uh this includes
0:01:10	to define
0:01:11	the divergence is
0:01:12	uh the proposed kernel
0:01:14	and then i got a little bit about exponential family nodded to see that to show that
0:01:18	at least for this a family distribution there's
0:01:22	i think not to risk
0:01:23	not heuristic involved
0:01:25	okay
0:01:26	so um
0:01:29	so we have this
0:01:30	well basically is that non parametric approach to plaster
0:01:34	uh the number of cluster a is not required known a priori
0:01:37	which means that it fit well to the problem
0:01:40	okay
0:01:40	it's awful mimosa units
0:01:42	should be considered rather after develop a hierarchical clustering
0:01:46	all these approaches
0:01:47	other kind of vacation or
0:01:49	yeah
0:01:49	this
0:01:50	i basically much segmentation
0:01:51	stuff like that
0:01:53	but also object right
0:01:54	recent
0:01:56	and my my reference is it
0:01:58	the seminal paper of uh combination read by me
0:02:02	uh i recently seconded
0:02:04	something like
0:02:05	citations are something like
0:02:06	three thousand
0:02:07	so it
0:02:08	similar did
0:02:10	uh some examples from the the paper you have any mention you wanted to segmented
0:02:15	based on the colours
0:02:16	okay and you have this
0:02:17	eh
0:02:18	and at one parameter that something so you never mind
0:02:21	uh
0:02:22	your target it is to define these clusters are you see that
0:02:26	hi debbie dorsey
0:02:27	so
0:02:28	that's the reason why we go no permit
0:02:31	if you wanna
0:02:31	one or a find a parameter model you
0:02:35	it would be a disaster
0:02:36	see this
0:02:37	also
0:02:38	something like this you see them all
0:02:40	there are seven models
0:02:42	if you want to have a gaussian distribution then be considered is that it
0:02:46	with not with a if you prevail probably doing
0:02:50	so another example of the new position how the original about here
0:02:54	and
0:02:55	uh this is a very the call or a
0:02:59	bandwidth salmon explained and this also is
0:03:02	this possible
0:03:04	how smooth you wanna be and
0:03:06	but this way you
0:03:07	have different levels of smoothing
0:03:09	okay this is another example uh again
0:03:12	paper
0:03:14	very similar approaches
0:03:15	you want to extract the boundaries
0:03:17	so what do you do a colour segmentation first and then you have to see what
0:03:21	there is about the same here
0:03:23	so
0:03:25	the limitations now
0:03:26	in order to uh the data they don't to
0:03:29	what like
0:03:30	directly to our problem
0:03:32	well that's on the spatial of observation
0:03:34	see the text
0:03:35	you know parameterisation
0:03:37	okay whereas uh uh there are several uh class and task
0:03:40	like
0:03:41	the one we have here
0:03:43	where
0:03:43	the natural end
0:03:45	please
0:03:45	we are we have can only be described using
0:03:48	parametric models
0:03:51	so
0:03:52	can we adapted
0:03:53	in order to be applicable for problems such problems
0:03:55	they're not they're missing the problem so
0:03:57	so i have some photographs
0:04:00	with different analysis
0:04:01	and we want to plaster
0:04:03	unsupervised
0:04:04	the same problem
0:04:06	you wanna have described it saved by um a normal distribution
0:04:09	each of the four
0:04:11	and you wanna
0:04:12	if you want to live in this missive algorithm in order to classify class of them
0:04:17	you wanted to do this model the observation space with
0:04:20	there's not euclidean geometry which is in euclidean geometry let's say
0:04:24	but
0:04:25	a lot of space
0:04:26	so
0:04:27	the proposed method i suppose used as an exponential family uses
0:04:30	a bayesian framework
0:04:32	and they just some concept
0:04:33	or information john
0:04:35	some
0:04:39	so
0:04:40	standard is
0:04:41	you know parameter
0:04:42	estimation
0:04:43	you have some data by X
0:04:44	this X matrix
0:04:46	and using a possible window in order
0:04:48	six
0:04:48	smooth
0:04:49	the empirical distribution you convolved with kernel
0:04:53	let's say see possibly still have is rather close to mitigate no
0:04:57	say put it's a gaussian
0:04:58	and
0:04:59	the only parameter here is
0:05:00	V eights about
0:05:03	excel
0:05:04	something okay
0:05:06	department impersonal parameter recovered that stuff before
0:05:09	but
0:05:10	what what not grammatically basically means that you let you parameters grow linearly with your data
0:05:15	doesn't mean that you don't have parameters actually
0:05:17	but you're parameters are actually the data themselves
0:05:20	and
0:05:20	last someone with some say stuff like that
0:05:23	okay and the buttons can be bible too
0:05:26	um the basic problems are mainly that you don't have enough data to estimate
0:05:31	properly
0:05:32	okay and that's to have more more a
0:05:34	uh
0:05:35	the dimensions
0:05:37	um you require more more data
0:05:39	i guess of the personality and all the stuff
0:05:44	but
0:05:44	the point is
0:05:45	do we actually need
0:05:46	to to
0:05:47	to estimate robust
0:05:49	for each problem
0:05:51	these on the like yeah
0:05:52	and the answer is no for example
0:05:54	class and
0:05:56	consider before the before
0:05:58	what we need it you wire is
0:06:00	a method
0:06:01	because
0:06:02	because the mode
0:06:03	say they have said
0:06:04	we haven't we should have another but not the the become the mode
0:06:08	and the method
0:06:09	plus sign each observations
0:06:11	to the appropriate mode
0:06:12	whatever this means
0:06:13	so
0:06:14	should we require the estimate
0:06:16	robustly that we get
0:06:17	oh that's really
0:06:18	if we can easily by
0:06:20	bypass
0:06:21	this procedure
0:06:22	and that
0:06:23	what is it that
0:06:24	so recall the expression
0:06:27	find more
0:06:28	seem to differentiate
0:06:29	with respect to X
0:06:30	and set this to zero
0:06:33	okay
0:06:34	after some ads in because this is a square distance you squarely so you have this you know
0:06:39	so
0:06:40	define for it
0:06:41	that
0:06:42	the differential of the current no
0:06:44	with these G
0:06:45	and you have the simple form
0:06:47	apart from the constant you have this and this this is
0:06:50	you can interpret it as
0:06:52	uh say the estimated pdf using
0:06:55	the differential profile
0:06:57	and the other is a message that
0:06:58	the main the main a result
0:07:00	this
0:07:01	what
0:07:02	let's say is is that
0:07:03	it reminded
0:07:05	okay when
0:07:07	you have this
0:07:08	uh say wait time range
0:07:10	of all the pixels
0:07:12	with respect the differential kernel
0:07:15	when
0:07:16	and and you all you need to do is
0:07:18	find a way to buy this
0:07:21	if
0:07:21	this is one of the main means that you are you know more
0:07:25	it so you don't actually estimated dense
0:07:27	you seem to do this
0:07:28	and that's implies the other
0:07:30	so that would look like this
0:07:32	very intuitive
0:07:34	this is abusive spectrum
0:07:36	so for each of the reservations
0:07:38	and the other does not matter at all
0:07:40	start begin with excited to to say X zero
0:07:44	calculate an assist vector
0:07:46	like this
0:07:48	okay
0:07:49	and simply
0:07:50	and it
0:07:50	that one position
0:07:52	until convergence it they are proven over this is very
0:07:56	trivial insulting the common need to set up a
0:07:58	so that was it
0:07:59	due to the situation for all
0:08:01	the observation huh
0:08:03	okay and stored into the observation here
0:08:06	the convergent point
0:08:07	so
0:08:09	if
0:08:10	two
0:08:11	or more are on the same old
0:08:13	they belong to the same cluster
0:08:15	does it was that
0:08:16	so
0:08:16	if you wanna
0:08:17	the
0:08:18	post here is an observation
0:08:20	okay the initial position
0:08:21	you see the trajectory
0:08:22	any
0:08:23	but
0:08:23	those are
0:08:25	for the red the red apple
0:08:28	thus it was that that's um is it ugly
0:08:30	the main idea
0:08:31	in observation space
0:08:33	okay
0:08:33	and you can and class was used in arbitrary shapes
0:08:38	see we're clustering
0:08:41	what are made with a gun
0:08:42	and
0:08:42	uh
0:08:44	make this last week
0:08:45	be one i mean
0:08:46	it's very very good
0:08:48	so
0:08:50	yeah
0:08:55	so how how can we had that this idea
0:08:57	be applicable to the spatial distribution
0:09:00	suppose they could be last on the same have
0:09:02	and utterances
0:09:04	okay
0:09:05	um
0:09:07	not forget about the speaker last numbers are taken more generally have an distributions
0:09:11	parameterised by theta
0:09:13	we should if i colonel
0:09:15	that means a shape in the distance
0:09:16	appropriate
0:09:18	okay and the pdf can be regarded
0:09:20	as opposed to a few time which sense
0:09:22	in which is that you consider
0:09:24	the density of the data
0:09:26	given
0:09:27	your observations
0:09:28	and this is it i would
0:09:29	what determines
0:09:31	or simply the cluster indicators
0:09:33	oh
0:09:34	your initial segmentation
0:09:35	so if you have considered that
0:09:37	speaker clustering task in that or is it some more
0:09:39	you apply first
0:09:41	segmentation it might be uniform might be
0:09:44	speakers said
0:09:45	based on speaker change detector
0:09:46	and this is it
0:09:48	okay
0:09:49	so in this sense as opposed to fit
0:09:52	and here is an example
0:09:54	suppose we have six
0:09:55	in this of segments
0:09:57	okay
0:09:58	and
0:09:59	this
0:09:59	yeah
0:10:00	the common to get the same when this is over all posterior
0:10:02	so if one apply the same idea and they begin with here
0:10:06	we will see that this last
0:10:08	would be attracted
0:10:09	only by itself
0:10:10	so we would create
0:10:11	a second bite so
0:10:13	okay the same haven't so this
0:10:16	the other three
0:10:17	we're gonna have that would be i don't think that we're gonna have
0:10:20	four
0:10:20	iconoclast
0:10:22	and the other the
0:10:23	the the last one again
0:10:25	its own class
0:10:27	see this again to the remote files but it's it
0:10:30	exactly the same idea
0:10:33	okay so it's
0:10:34	like
0:10:35	a higher level in the hierarchy that's what
0:10:38	that's
0:10:38	a integrated to what if i have the observations
0:10:42	and the parameters but now
0:10:43	you are in the space of observations
0:10:45	and you have a posterior
0:10:47	okay and uh
0:10:49	the the same way you want express somehow the uncertainty in a smoother results
0:10:53	by using this kernel
0:10:54	in the observation domain which might be gaussian
0:10:57	on the same way you have to be
0:10:58	you have to expose your uncertainty
0:11:00	about uh the estimation
0:11:02	and as you see like
0:11:05	why why we should a consider also
0:11:08	the the the number
0:11:10	the sample size of its class it was supposed to have the same position
0:11:14	and supple somehow
0:11:16	uh they were
0:11:17	all this corresponded to ten times
0:11:19	that's the sample size
0:11:20	before
0:11:21	well
0:11:22	then probably all all these classes
0:11:24	would be single tones
0:11:26	it's saying that would be it would be a single class because we expect
0:11:29	that as much as more data right
0:11:32	these three
0:11:33	we will manage
0:11:34	more to what we would have a close
0:11:38	okay so
0:11:39	there is certainly dependence of the sample size if it's linear
0:11:42	it's not
0:11:43	i i
0:11:43	i
0:11:44	it's a linear or only
0:11:45	if the motors are correctly specified
0:11:48	and in speaker diarization and especially for using
0:11:51	simulation model is
0:11:52	there's if you just dismiss misspecification
0:11:56	that
0:11:56	you can consider this
0:11:57	actually
0:11:58	and that's a problem
0:12:00	um so
0:12:01	let's define the kernel
0:12:03	um
0:12:05	let's see this delegation this some months
0:12:07	but uh probably don't have time line
0:12:09	but
0:12:10	consider this as a parameterised by delta
0:12:13	family of the aviation be to assume so
0:12:15	the for me to scale endeavours
0:12:18	and the others
0:12:19	but you have living their argument
0:12:21	yeah that was the liquid
0:12:23	you okay
0:12:24	uh they wanted to was what the hell at least
0:12:28	and is your estimate distance
0:12:30	but you can also estimate right and all that kind of errors by summing it
0:12:33	or by taking their harmonic mean
0:12:35	and
0:12:36	we use this approach
0:12:37	however many dressing up was um
0:12:39	can based on this highly reduced
0:12:41	okay and recall that
0:12:43	this will happen no matter what the database
0:12:46	which is a
0:12:47	this is the see that pitch information about this which is simply that
0:12:50	then they made the metric tensor
0:12:52	you do it you could consider it information john
0:12:57	so having defined as the shape
0:12:59	now consider a some infinite distances let's define the shape
0:13:02	because
0:13:03	we had to go motion
0:13:04	in the observation space
0:13:05	should be goes any longer
0:13:06	well
0:13:07	it's another parameterisation now by myself
0:13:10	okay if you if you consider only out of that
0:13:13	are equal to one
0:13:15	then you have an exponential
0:13:16	okay and this should be considered
0:13:18	as i read the rather nicotine
0:13:20	the K derivative
0:13:22	density with respect to the information element in these information element with like real big measure if one
0:13:28	but anyway
0:13:29	um well
0:13:30	so
0:13:30	if you don't like
0:13:31	uh though i i i
0:13:33	simply consider a close one
0:13:35	you have a lot of time for mobile phone
0:13:38	and isaacs play a video propose again explain you collected yesterday
0:13:41	these are varied
0:13:42	only two
0:13:43	the T distribution
0:13:45	if you wanna
0:13:45	considering
0:13:47	us a heavy tail
0:13:49	priors
0:13:50	okay so it's a very nice interpretation of that
0:13:52	the looks like so
0:13:54	i'm not gonna
0:13:55	analyse all this stuff
0:13:57	but all this
0:13:58	and we can see that by minimising the cost function
0:14:01	okay here
0:14:03	this down
0:14:04	is
0:14:05	correspond to how much does you are off in the you are about to measurement
0:14:09	so
0:14:10	it should be somehow leno leno
0:14:12	with the sample size
0:14:14	whereas this to tell you
0:14:16	how close
0:14:17	you wanna be
0:14:18	with an informative prior the jeffrey's prior
0:14:20	basically
0:14:21	which is simply the flat prior if you consider
0:14:24	uh you can enjoyment
0:14:26	i think does that so its minimisation of cost function
0:14:29	and here is
0:14:29	just a single the average
0:14:31	'kay on them
0:14:31	yeah that
0:14:32	then the deviation
0:14:34	okay
0:14:35	that's great and
0:14:36	all this
0:14:37	uh right of the conjugate priors all this stuff
0:14:40	and we consider our as um
0:14:43	parameter parameter additional this family
0:14:46	but it's often end up
0:14:49	so
0:14:50	let's go back problem having to find all the stuff
0:14:53	so here's the posterior
0:14:55	we have to have okay segments
0:14:57	okay
0:14:58	this is simply the difference but i don't
0:15:00	you can be outside
0:15:02	and you have these
0:15:02	suppose it
0:15:03	i would like also one let's consider only the stuff
0:15:06	okay
0:15:06	so this should be normally something like away
0:15:09	and K
0:15:10	are simply the number of my examples of how of it for the case uh sec
0:15:15	and then a role
0:15:16	seamless are just like like a weight so you may consider doesn't it
0:15:19	as a as a mixture of uh distribution
0:15:22	okay so the point is now
0:15:25	recall that have differentiated respect to extract because of
0:15:28	of
0:15:29	uh using the squared distances
0:15:31	we came up with this formula
0:15:32	so the question is
0:15:34	we did we would you heuristics and a or a by differentiate this stuff
0:15:39	or or or not
0:15:41	uh and and the point is that
0:15:43	if you constrain yourself
0:15:45	you'd exponential family
0:15:47	then there's not reasonable for example in the in the
0:15:52	save for the normal distribution
0:15:54	you have these uh the this a natural parameters that correspond to this problem tradition
0:15:59	the sufficient statistic that
0:16:00	i a tax naturally with the the
0:16:03	the natural parameters
0:16:04	and the measurements constant
0:16:06	and um but i'm not gonna get into that because
0:16:09	the wine tasting wait
0:16:11	um
0:16:13	so
0:16:15	there is another couple
0:16:17	uh
0:16:19	or a parameter space
0:16:20	and is the expectation parameters
0:16:22	more close to what
0:16:24	we usually
0:16:25	uh tree
0:16:27	in this
0:16:27	in in the speaker clustering
0:16:29	okay
0:16:30	by exploiting
0:16:31	the complexity
0:16:33	or the log partition with respect
0:16:34	to defeat that
0:16:35	you end up with a not with the expectation parameters
0:16:39	okay and
0:16:40	by
0:16:41	by re differentiating the stuff
0:16:43	with respect to feed that
0:16:45	you get
0:16:46	the fisher information
0:16:47	major
0:16:48	matrix okay
0:16:50	all the stuff or
0:16:51	from the classical statistics of where known
0:16:54	and
0:16:55	this is a potential find that the two potential functions
0:16:58	which shows the mostly clearly why this
0:17:01	ah
0:17:02	this pay some couple expectation parameters
0:17:04	and that he the parameters
0:17:06	so if you want to do what you wanna do what you want to read
0:17:10	and the same way you average between points in there
0:17:13	euclidean domain
0:17:15	you may consider also
0:17:17	the
0:17:17	expectation parameterisation
0:17:19	the expectation parameters
0:17:21	all the natural parameters
0:17:22	and it depends on you
0:17:23	but like to do
0:17:25	okay which geometry you wanted to
0:17:27	do you probably have
0:17:29	which more appropriate
0:17:31	so
0:17:33	they could but lyman let's let's consider only the K and i went to the two extremes
0:17:38	the one and zero
0:17:40	they'll take one or zero
0:17:41	you have these
0:17:42	that's that's these are all mad
0:17:44	but
0:17:45	if you differentiate you obtain some cleaner
0:17:47	what to do this
0:17:48	but it also this
0:17:49	there's a there is any finally show like to mention here that's not mentioned my paper
0:17:54	you expect that
0:17:55	by differentiating this and
0:17:57	and they are like rock and roll
0:18:00	you have a parameterisation with that data
0:18:03	but
0:18:04	since you don't use the natural gradient
0:18:07	okay see that sees you don't consider the care
0:18:10	but this new space is curved
0:18:12	you differentiate respectively attended in if you switch composition
0:18:17	which is somehow weird
0:18:18	but if you use the natural gradient
0:18:21	which define simply defined as this
0:18:23	with distilled that
0:18:25	okay
0:18:26	now if you differentiate with respect to feed that
0:18:29	you remain to the same parameterisation
0:18:31	okay there's no switching
0:18:33	between this parameterisation
0:18:34	so you know i recall that in order to to make it a
0:18:38	to define a ground rule
0:18:40	you should
0:18:40	want to do the kind of the space
0:18:42	so you should
0:18:43	work with the natural reading
0:18:47	so
0:18:48	the final
0:18:50	'cause i did what you put in like a month in the model this model called
0:18:54	okay
0:18:54	recalling the same again
0:18:56	iteration
0:18:57	start from a from the percent per segment of the last there's no doubt that has no matter
0:19:02	but how it doesn't matter
0:19:04	and your next estimation will be an average
0:19:07	oh for
0:19:09	of all the all for the other segment
0:19:11	in this parameterisation
0:19:12	or in this parameterisation the fit
0:19:15	they're not around
0:19:16	it depends on you
0:19:20	so
0:19:21	suppose we have some segments
0:19:23	it's all segments
0:19:25	and the method to sixteen clusters
0:19:27	it's blue uh don't
0:19:29	represents
0:19:30	actually the mean the mean value
0:19:34	okay the mean value of these say the these say to mfcc
0:19:39	visions
0:19:39	okay and see how the man
0:19:41	you know sixteen plus you also find
0:19:43	some single class
0:19:45	that simply that right
0:19:47	by the bottom so
0:19:51	so
0:19:52	something a final report to go to the experiment
0:19:56	um uh
0:19:58	all this i wouldn't feel about the U two specification is no
0:20:02	uh you should somehow bias
0:20:04	the results
0:20:05	um towards that have an annuity because you know
0:20:09	that a a dialogue with the doesn't have
0:20:12	say utterance
0:20:13	of say three or four a
0:20:16	segments
0:20:17	is this a must buy as a as a a supplying a a
0:20:21	do you play prior
0:20:22	over a transition matrix
0:20:24	been forced somehow continue eating
0:20:26	so
0:20:27	but we do
0:20:28	we multiply
0:20:30	okay
0:20:30	by this uh C distribution
0:20:34	okay so if you are
0:20:36	supporting the in the cave
0:20:38	segment
0:20:39	okay
0:20:39	you want to emphasise your neighbour
0:20:42	but not in a way that goes and does it
0:20:45	okay not in such a way because if you if you do together some way
0:20:49	the first and the last segment would we what we want one B one from the same a classroom
0:20:55	never
0:20:55	you wanted it more mild and everything
0:20:58	yeah i guess
0:20:59	i don't the one week
0:21:01	so
0:21:02	is there a database
0:21:05	okay
0:21:07	sure
0:21:07	oh well
0:21:08	you know this is a very database broke
0:21:10	news
0:21:12	okay and we compare it to the standard of big uh here approach
0:21:16	you won't find any and super duper while our results
0:21:20	uh
0:21:21	the best
0:21:22	a configuration we find with local because this
0:21:25	we fixed with fixed the lab agenda parameter using the development and then we have this
0:21:29	result in the test set
0:21:31	and M S was put them in C
0:21:33	exactly
0:21:34	hey using only single ego self motivated matrix
0:21:37	no gmms at all
0:21:38	and we see that the harmonic mean was the best
0:21:42	four
0:21:42	and the other side
0:21:43	rather close enough not though
0:21:46	if you use
0:21:47	and if you're a clustering
0:21:48	seem to care that it is
0:21:50	would be a tragedy
0:21:53	this is because of the big
0:21:54	right you
0:21:55	but you cannot use big you know
0:21:58	and that's a problem
0:21:59	i would use this stuff
0:22:00	so
0:22:02	to finish
0:22:03	but also
0:22:05	and adaptation of an open out of attacks on the space of observation space bounded
0:22:10	okay we use
0:22:12	some reasonable i think uh bayesian argument in order to
0:22:16	to make these transition
0:22:19	okay and i was so that at least for the exponential families
0:22:22	not recyclable and stuff
0:22:24	so
0:22:25	uh well these are relevant to be honest um
0:22:27	you want to obtain a point estimate about your hidden variables
0:22:31	build stuff document clustering but
0:22:34	or if you wanna have diarrhoea that
0:22:36	you
0:22:37	consider all real time lapse
0:22:39	but if you wanna doing force you don't you
0:22:41	stuff
0:22:42	you don't use your clustering either
0:22:44	okay you do do do do dirichlet process
0:22:46	says
0:22:47	you use um N C M C
0:22:50	you use variational bayes
0:22:51	okay
0:22:52	so
0:22:53	one final
0:22:54	gmms
0:22:55	certainly you can if you consider the complete data they don't if you consider
0:22:59	they as a like the complete data likelihood
0:23:02	then the gmms we don't wanna see
0:23:03	family
0:23:04	yeah
0:23:05	but you you need to have a correspondence or you need to start
0:23:08	with the ubm with a common ubm
0:23:10	why because
0:23:11	if you can see the the complete data likelihood
0:23:14	you only
0:23:15	i can see there
0:23:16	uh okay on the basis between
0:23:19	corresponding
0:23:20	uh gaussian
0:23:22	last
0:23:23	italian
0:23:24	for them different weights if you
0:23:26	trained if you also allow the trains
0:23:29	the the weights to be to be
0:23:31	also i vectors
0:23:32	i'm sure you can use it for
0:23:34	that
0:23:35	even the original message
0:23:36	however
0:23:37	if you use
0:23:38	i i picked those
0:23:40	note that you have
0:23:41	um different segments
0:23:44	so
0:23:45	the variability of the estimate
0:23:46	should also be part of the problem
0:23:49	so somehow
0:23:50	um
0:23:52	play with a band with allow the bandwidth to be
0:23:54	depending on the sample size
0:23:56	okay to to to encode these
0:23:58	i uncertainty
0:24:00	in the this estimate
0:24:03	thank you
0:24:10	thanks and sorry for
0:24:11	chris you wanted to keep some
0:24:13	time
0:24:14	for some question
0:24:15	ooh
0:24:16	you should comment
0:24:24	fig
0:24:25	okay
0:24:26	yeah
0:24:26	quick
0:24:27	just simple
0:24:29	it's on
0:24:30	it's one
0:24:30	uh
0:24:31	you are optimising
0:24:33	uh
0:24:33	you time for them
0:24:35	yeah
0:24:35	so
0:24:36	where
0:24:37	doing gradient descent
0:24:38	so
0:24:39	you can
0:24:40	you have to compute the gradient
0:24:42	but
0:24:43	do i understand correctly that you're not able to compute the
0:24:47	value
0:24:48	so you during the optimisation
0:24:49	oh
0:24:50	being
0:24:51	i don't evaluate don't it actually you can this is the
0:24:55	the house analytical that
0:24:56	but is not correct
0:24:58	it's based on that
0:24:59	you know it's uh
0:25:01	it's a simple estimate
0:25:03	what what what
0:25:04	the idea behind miss that means if i was it
0:25:06	that you don't actually need
0:25:07	to estimate the overall period you can bypass problem by using gradient
0:25:12	that the idea
0:25:13	is the
0:25:14	just a
0:25:15	practical thing because my favourite optimisation
0:25:18	yeah
0:25:20	right objective function
0:25:22	okay
0:25:23	if we go
0:25:24	okay
0:25:25	so
0:25:26	of the wine
0:25:31	well we should come in
0:25:41	general questions about the mean shift
0:25:44	um
0:25:45	if you start off
0:25:46	start out by hypothesising time clusters
0:25:50	you will always got time clusters is that correct
0:25:53	no you don't no or little number one stress come you
0:25:57	some
0:25:58	i was just thinking proper
0:26:00	the number of classes completely fair infer from the from the average
0:26:04	you don't you don't the uh
0:26:07	proof a predefined number of class
0:26:10	it depends only if the if the points compared
0:26:13	okay to the to the same value
0:26:15	so
0:26:16	you don't need to
0:26:18	big
0:26:18	yeah
0:26:19	you said that um
0:26:22	not having not being able to incorporate
0:26:24	a course
0:26:25	first of all you cannot have big because
0:26:27	big big implies
0:26:29	a a marginalisation with respect to the parameters
0:26:33	okay but it isn't it it's a transfers dropout
0:26:37	naturally hmmm it's
0:26:39	thrusters
0:26:40	dialling
0:26:41	not sure
0:26:43	you know you can hold
0:26:45	find the correct number of clusters
0:26:47	with that without using a big
0:26:49	sorry
0:26:49	okay
0:26:49	material
0:26:51	sure sure
0:26:52	i'm not using it because i forgot the comparison
0:26:54	probably the correctly last
0:26:59	okay

Speaker clustering via the mean shift algorithm

SESSION 8: Human performances in Speaker recognition, Speaker clustering and partitioning

Přidáno: 14. 7. 2010 11:08, Autor: Themos Stafylakis (Institute for Language and Speech Processing, National Technical University of Athens), Vassilis Katsouros (Institute for Language and Speech Processing), George Carayannis (National Technical University of Athens), Délka: 0:27:06