Přepis řeči - Unsupervised Speaker Adaptation based on the Cosine Similarity for Text-Independent Speaker Verification

0:00:06	more
0:00:07	really
0:00:08	okay good
0:00:10	right okay so
0:00:11	today i'm gonna start
0:00:12	talking a little bit about um unsupervised you adaptation
0:00:15	and uh
0:00:16	with respect to use a total variability cosine scoring like it changes
0:00:20	previously discussed but
0:00:22	first off
0:00:24	i've only known a gene for for a couple of months
0:00:26	and
0:00:27	and i think a lot of you probably know a lot more about an idea but
0:00:31	i just want to summarise maybe a couple of things
0:00:33	i where
0:00:34	and
0:00:35	and these are in in court said that he's had over the last couple of a
0:00:39	couple weeks now
0:00:40	um i i had to get a little bit appear approval with with with this 'cause i wasn't sure this
0:00:45	introduction slight was gonna be appropriate at all
0:00:47	but well
0:00:48	it seemed okay so we're gonna go with that
0:00:50	so this is what happened
0:00:52	during a
0:00:53	this sre do that okay
0:00:54	this is this is this main but
0:00:57	um
0:00:58	no
0:00:59	upon arrival in brno
0:01:01	um
0:01:03	it goes ahead
0:01:07	oh
0:01:08	yeah
0:01:11	but but i like his mind changes really quickly you 'cause
0:01:14	i mean
0:01:15	and not just put this together bottom
0:01:17	a couple hours ago 'cause like this morning he was really
0:01:20	it's really up for the city is
0:01:21	is good
0:01:22	fig
0:01:24	you know
0:01:25	no words that and so i decided you know protect what what's gonna happen in a few days and like
0:01:32	and and this is kind of what i'm what i'm hoping for
0:01:34	as you as you may know
0:01:36	he doesn't
0:01:36	during
0:01:37	that much
0:01:38	here anything but i'm thinking that maybe
0:01:41	he also for co for some reason um but
0:01:44	we maybe we should
0:01:45	we should help them out with with fear and an open on
0:01:48	give that a go anyway
0:01:49	down to business
0:01:50	um
0:01:52	so
0:01:54	the the whole idea of how my top
0:01:56	is that
0:01:57	uh it's on
0:01:58	again
0:01:58	unsupervised
0:02:00	adaptation and and the whole motivation behind it is that uh
0:02:03	capturing characterising every source
0:02:05	oh variability
0:02:06	um is pretty difficult especially with only one enrolment session
0:02:10	now
0:02:11	if we were
0:02:12	able to have multiple enrolments of the same speaker
0:02:15	this would
0:02:15	help
0:02:16	average out
0:02:17	um more sources of the inconsistencies
0:02:20	and provide a better representation
0:02:22	apart
0:02:22	and a speaker model
0:02:24	and so
0:02:25	this brings us to the problem of um
0:02:28	so you can adaptation in general
0:02:30	um
0:02:30	but
0:02:31	in the problem unsupervised speaker adaptation we are updating our speaker models with out
0:02:35	a priori knowledge
0:02:37	that utterance that were
0:02:38	updating a model with
0:02:40	actually belongs to the target speaker
0:02:42	now
0:02:43	and we do so based on utterance
0:02:45	process during testing and this
0:02:47	is
0:02:47	was incorporated in uh
0:02:49	i think in this very uh two thousand four
0:02:51	what isn't five or so
0:02:53	um
0:02:54	now
0:02:56	in previous work
0:02:57	um using joint factor analysis
0:02:59	before
0:03:00	we began on it
0:03:01	work with total variability
0:03:03	um what we noticed was that there were indeed
0:03:05	highly variable scores
0:03:06	on that were produced
0:03:07	by J F K
0:03:08	and the required normalisation in particular is each you know or
0:03:12	um
0:03:14	and applying
0:03:15	these
0:03:15	score normalisations
0:03:17	in the unsupervised adaptation domain requires
0:03:19	a significant amount of additional computation
0:03:21	with each adaptation of
0:03:24	um i'll go a little bit more into that
0:03:26	um just a little bit
0:03:28	now
0:03:29	when we when i
0:03:30	we begin this attend
0:03:32	with total variability
0:03:33	we were hoping for a certain number
0:03:36	of improvement
0:03:37	where
0:03:38	um we could
0:03:39	do
0:03:40	unsupervised adaptation with
0:03:42	less computation
0:03:43	um and we take advantage of
0:03:45	total variabilities
0:03:46	use of lowdimensional total factor vectors or ivectors
0:03:50	we can debate on who once
0:03:51	to name what later
0:03:53	um but
0:03:54	and then there's a cosine similarity scoring which is
0:03:57	which is very que
0:03:58	and
0:03:59	and
0:04:00	also there's also the news
0:04:02	a set of news
0:04:02	score normalisation
0:04:04	strategies that we wanted to play with
0:04:06	namely the the symmetric normalisation snr
0:04:08	and uh
0:04:09	normalised cosine disk
0:04:10	that is you just talk
0:04:13	so
0:04:14	the a little bit about sort of a
0:04:16	outline for this talk on
0:04:17	i i'm gonna
0:04:18	go over
0:04:20	really quickly whatever
0:04:21	total variability
0:04:22	wasn't as in just
0:04:23	are you get a very good job explaining
0:04:25	some of on some of the ideas behind it
0:04:26	oh going to uh
0:04:27	the the unsupervised adaptation
0:04:30	algorithm that that we
0:04:32	then we came up with on that that's
0:04:34	gotten decent
0:04:35	is that
0:04:35	in in results and then
0:04:37	we can um
0:04:38	proceed onward with uh
0:04:40	score normalisation experiments
0:04:42	and a little better for the disk
0:04:47	so total variability et cetera
0:04:49	has
0:04:49	as all the components
0:04:51	that um
0:04:52	that was shown
0:04:53	in the previous thought
0:04:54	and and that we've
0:04:55	probably
0:04:56	all seen
0:04:57	in the past
0:04:58	on this
0:04:59	you know we're using factor analysis
0:05:00	feature extractor you have a
0:05:02	speaker and channel dependent supervector
0:05:04	um
0:05:05	there's intersession compensation without the in the V C C N
0:05:09	and cosine scoring
0:05:10	um
0:05:11	just
0:05:12	in the final at the end of the day we're just gonna use um
0:05:15	W prime
0:05:16	um
0:05:17	which is after everything has been applied
0:05:19	and such that was an scoring is actually really just
0:05:21	um the you know product between that you've
0:05:26	so
0:05:28	for previous work
0:05:29	in
0:05:30	in joint factor analysis
0:05:31	um
0:05:32	on the topic
0:05:33	of unsupervised adaptation
0:05:34	was
0:05:35	done by dan rosen kenny
0:05:37	um
0:05:38	and what they did
0:05:39	was given some new adaptation data they would
0:05:42	compute
0:05:43	the posterior distribution of that speakerdependent hyperparameters
0:05:46	using um the current
0:05:47	ever
0:05:47	odours
0:05:48	as
0:05:48	prior
0:05:49	um
0:05:50	and what they what they also did was to set a fixed and predefined
0:05:54	adaptation threshold
0:05:55	and use log likelihood ratio scoring
0:05:58	um
0:05:59	the in that paper they also introduce an adaptive
0:06:02	T known score normalisation technique
0:06:04	um because what they had observed
0:06:06	was there is a there's a good
0:06:07	this attribute you normalise scores as more adaptation data was used
0:06:11	um and
0:06:12	and
0:06:13	in order
0:06:14	two
0:06:14	use
0:06:15	um be able to use
0:06:16	fixed decision threshold
0:06:18	with
0:06:19	um in their in their decision process
0:06:21	um they had to do
0:06:23	a a new type of normalisation
0:06:25	now
0:06:26	that that that that was met with
0:06:27	a good amount of success
0:06:29	and um
0:06:30	and they were very promising yet
0:06:32	that to say in order to combat
0:06:33	the
0:06:35	in order to implement the adaptive you know normalisation it requires a good bit of computation
0:06:41	um
0:06:41	and
0:06:42	um calculating pursue to speech design easy
0:06:45	and um
0:06:47	and also there was
0:06:49	that there was a require a computational as you know in france
0:06:52	every
0:06:53	cation update
0:06:54	um so
0:06:56	uh and then lastly i guess
0:06:57	success also dependent on the choice of adaptation threshold which was
0:07:01	which was tuned but
0:07:02	well we also get an animal
0:07:03	but
0:07:04	now
0:07:05	for us then
0:07:06	in order
0:07:07	two
0:07:08	better
0:07:09	uh or try to improve upon this work
0:07:11	in in the context
0:07:12	of total variability
0:07:14	um what we wanted
0:07:15	was
0:07:15	satisfy the following criteria
0:07:17	we wanted a simple and robust method for setting an adaptation threshold
0:07:21	data
0:07:22	and what we decided would
0:07:23	was
0:07:24	to set it to be insane
0:07:25	and some optimal decision thresholds and development data
0:07:28	but we're gonna use was the nist two thousand
0:07:30	six
0:07:30	um
0:07:31	sre
0:07:32	sre data and
0:07:34	basically
0:07:35	um
0:07:36	what we would do
0:07:37	uh is
0:07:40	carry out your test without adaptation
0:07:42	um
0:07:42	are
0:07:43	carry out of one without adaptation and set
0:07:45	the optimal
0:07:47	i guess the the the point that will
0:07:49	minimise the dcf
0:07:51	i i a threshold and we'd set that as a special to to test
0:07:54	um i don't get in to details about that in a little bit
0:07:57	no
0:07:58	next was
0:07:59	we wanted
0:07:59	minimise
0:08:00	the amount of computation that was that's
0:08:02	area during each
0:08:03	unsupervised adaptation update
0:08:05	and
0:08:05	this helps that already in
0:08:07	control variability
0:08:09	we are we are able to use low dimensional
0:08:11	um
0:08:12	total factor vectors
0:08:13	and that we are able to use
0:08:14	cosine similarity scoring
0:08:16	and lastly
0:08:17	our hope was
0:08:18	to simplify
0:08:19	score normalisation procedures wherever part
0:08:25	so
0:08:25	really the really basically i mean if you were
0:08:27	if we're able to use our total factor vectors or i vector
0:08:30	as
0:08:31	um
0:08:32	when estimates
0:08:33	um
0:08:33	in in our speaker space then
0:08:35	then given a limited training data we we might not obviously have
0:08:39	it's perfect
0:08:41	estimation of
0:08:42	of where our speaker really lies
0:08:44	and so i suppose
0:08:46	this is just the cartoon so it's
0:08:47	it's not
0:08:48	not anything rigorous at all
0:08:50	but suppose we had our
0:08:51	true speaker identity S which is
0:08:53	um which can be given
0:08:55	by
0:08:55	by the
0:08:56	the little circle there
0:08:58	but then
0:08:58	and are estimated
0:09:00	and one utterance or speaker identity of um the reason i
0:09:03	and so
0:09:04	this might not be
0:09:05	this isn't exactly very
0:09:07	right on the spot where the true speaker identity is
0:09:10	but if we had a good number of these utterances i it would make sense that
0:09:13	um
0:09:14	we should
0:09:15	converges towards a better representation of
0:09:17	speaker
0:09:18	now was this
0:09:19	this question is also assuming a priori that the additional data that we actually have
0:09:23	is
0:09:23	from
0:09:24	on speaker S
0:09:26	now
0:09:27	um
0:09:28	so
0:09:30	as such
0:09:30	we
0:09:31	decide to propose this
0:09:32	on this algorithm here
0:09:34	and
0:09:35	in an effort to match
0:09:36	the
0:09:37	the the technical rigour of the previous two presentations
0:09:40	for me i decided to pack as much math
0:09:42	i could in this
0:09:43	in this one slide
0:09:44	um
0:09:45	but
0:09:45	basically
0:09:47	it we're saying that we have a set of total factor vectors that are
0:09:50	um assumed to pertain to an identity
0:09:52	of a known speaker S
0:09:54	um
0:09:55	and then we have a set
0:09:56	of total factor vectors
0:09:58	T survive
0:09:59	on that are extracted
0:10:00	test utterance
0:10:01	each
0:10:02	of arc a test
0:10:02	and
0:10:03	with a defined if
0:10:04	decision threshold
0:10:05	data
0:10:06	um we have
0:10:08	a new equation for the score which
0:10:10	um
0:10:11	since
0:10:12	this is uh the
0:10:13	this notation is just the cardinality so that's is it just the mean of all possible
0:10:17	um and then we compared to some threshold and if the threshold
0:10:20	of of all the
0:10:22	and that threshold of the score exceeds
0:10:25	you're you're threshold then you decide
0:10:27	that you're
0:10:28	that
0:10:29	but the utterance
0:10:30	current utterance to supply would belong in inside here
0:10:33	into the
0:10:33	the identity of a new speaker as
0:10:36	and
0:10:36	you would
0:10:38	a you would
0:10:38	say yes to that trial and be you would
0:10:41	and the
0:10:41	um
0:10:43	that that new utterance these to buy into you
0:10:45	speaker big W sub
0:10:48	now
0:10:49	um what we have is the symmetry that allows for text
0:10:51	sure
0:10:52	and um later we will have
0:10:54	more discussion on the ideas
0:10:56	for on the design
0:10:58	of this function
0:10:59	is that it can
0:11:00	um conceivably be in a better
0:11:03	but
0:11:03	to to reiterate what i just said
0:11:05	um it's
0:11:06	it's actually
0:11:07	quite easy so if you had in
0:11:09	initial enrolment utterance
0:11:10	estimator speaker identity
0:11:12	on W
0:11:13	there is a one
0:11:14	and you have incoming test utterance to supply this is assuming that
0:11:17	these are all of your
0:11:18	speaker i i didn't is right it is all you have
0:11:21	just a single utterance
0:11:22	and that you compute a score than if you're
0:11:24	your singles what as one is great and data
0:11:27	then
0:11:27	you just
0:11:28	take
0:11:29	um you just take a test utterance and you can
0:11:31	place it
0:11:32	um
0:11:33	in with
0:11:33	with this and this becomes
0:11:35	what you have is now when you you test utterance
0:11:37	uh arcs are
0:11:38	and you training utterance
0:11:39	um
0:11:40	that that was
0:11:41	just to test
0:11:42	and so and that that's how you simply admit
0:11:45	um
0:11:45	a bit
0:11:46	more
0:11:47	training vectors into in your set
0:11:50	um and so now
0:11:51	you had a
0:11:52	second test
0:11:53	utterance then
0:11:54	and you can keep two scores
0:11:56	right you have
0:11:57	yeah you're initially
0:11:58	um estimated
0:12:00	speaker identity and this new
0:12:01	training i utterance W so uh T one
0:12:04	and you can see these two scores and now
0:12:06	if you um
0:12:07	if that function
0:12:08	of those two scores is
0:12:09	again greater than your fixed threshold data
0:12:11	then you do the same thing and
0:12:15	and do you
0:12:16	you put another
0:12:17	you training utterance
0:12:18	yeah
0:12:19	so
0:12:20	that's
0:12:20	that's all that's
0:12:21	and so
0:12:22	and the emphasis
0:12:24	behind this approach again is
0:12:26	such that we do not need to change decision thresholds data
0:12:29	um
0:12:30	previously that the idea i
0:12:32	in the past
0:12:32	in related work with that
0:12:34	with more
0:12:35	i i adapted utterances or
0:12:36	and all that in the text dependent setting
0:12:39	um
0:12:39	they
0:12:40	the
0:12:41	what some work was doing
0:12:43	in the past was just that you would um
0:12:45	increase
0:12:46	the decision threshold with each adaptation utterance before us
0:12:49	we want to keep things as simple as possible and so
0:12:51	we decided
0:12:52	that we did not wanna change in your decision there shop data
0:12:55	and also there's simply right now there's no modification of a troll factor vectors
0:12:59	simply all we're doing is actually combined score
0:13:02	um and now so that
0:13:04	summarises what variability and um unsupervised adaptation
0:13:08	something
0:13:08	go really quickly in the score normalisation which
0:13:11	i am aware
0:13:12	it's very well known topic
0:13:14	um
0:13:15	and so i'm not gonna give it a
0:13:17	give it a pretty
0:13:17	brief review
0:13:19	um with with uh
0:13:20	couple of
0:13:21	indices
0:13:22	um
0:13:22	in in the wording
0:13:24	but
0:13:27	so
0:13:27	it for the idea behind score normalisation
0:13:30	is that we are assuming that
0:13:31	distribution of target speaker
0:13:33	and impostor scores
0:13:34	followed to the string
0:13:36	two distinct
0:13:37	normal distribution
0:13:38	um and
0:13:39	however the the parameters for these two distributions are
0:13:42	on speaker independent
0:13:43	start part
0:13:44	target speaker depending as such we need to normalise to allow for
0:13:47	universal decision pressure
0:13:48	and so in zero normalisation ones enormous is well known
0:13:51	um we scale each of the distribution of scores
0:13:53	produced by target speaker
0:13:55	model and and a set of impostor utterances
0:13:58	two
0:13:58	standard normal distribution
0:14:00	now in test normalisation to no one
0:14:02	it's like the same thing it said
0:14:04	um
0:14:04	in order to adjust for the idea of intersession variability
0:14:07	we we
0:14:08	sh
0:14:09	scale into the distribution of scores produced by test
0:14:11	utterance
0:14:12	and the set of impostor models
0:14:14	um
0:14:15	to see a normal distribution now the the idea here is to keep
0:14:18	in mind that the ad tell size words um utterance and model
0:14:21	um not discuss how how the related in total but in the context of total variability in just a little
0:14:25	bit
0:14:26	is easy norm
0:14:27	we've already seen that um it achieves the best results
0:14:30	uh
0:14:31	in a factor analysis based system
0:14:33	um
0:14:34	and and that that's what's currently being used
0:14:36	stay there
0:14:38	so
0:14:38	now what we have
0:14:40	here with um
0:14:42	oh sorry uh what we have in G T not parameter updates during this model adaptation is
0:14:47	is
0:14:47	the lack of any
0:14:49	um
0:14:50	or
0:14:52	uh the need for
0:14:53	for more normalisation parameters right because
0:14:55	it is we can indeed become training utterances
0:14:58	um
0:14:59	W
0:15:00	the T one and diversity to then
0:15:02	on their previously test utterance
0:15:04	so they already had a T non associated with them
0:15:06	but
0:15:06	what we need additionally is obviously as you don't to be computed
0:15:09	but that's actually it so that means we can
0:15:12	simply
0:15:12	um precompute
0:15:14	roc norm parameters for each test utterance
0:15:16	as we do it
0:15:17	i
0:15:18	the same way we do it for each um
0:15:20	for each
0:15:21	target
0:15:21	figure utterances are
0:15:22	target speaker model as well
0:15:24	and so that's all we need to do
0:15:26	in in the past we had
0:15:27	compute is um
0:15:29	adapted to known parameters
0:15:31	um yeah that we can do is you don't parameters after each adaptation update however here
0:15:35	very simple and it's very in should be much quicker
0:15:38	um
0:15:39	now
0:15:40	the next
0:15:41	thing was
0:15:42	that
0:15:43	total variability
0:15:44	in the context of
0:15:45	the difference between
0:15:47	utterances
0:15:48	and
0:15:48	uh models
0:15:49	well
0:15:50	total variability uses factor analysis
0:15:52	the front end
0:15:53	and so
0:15:54	all we have is that
0:15:56	the extraction of total factors
0:15:58	from an enrolment
0:15:59	or test utterance
0:16:00	follows really the exact same process and as such
0:16:03	there's really no difference
0:16:05	between an utterance
0:16:07	and the model
0:16:08	and
0:16:09	and it there's no a and
0:16:11	also with the cosine similarity in the symmetry behind all that
0:16:14	um there is
0:16:15	no distinction to be made as that we can think of them all as the same thing
0:16:19	which brings us to an even more simplified method of score normalisation which is
0:16:23	which isn't which is the
0:16:24	S norm um
0:16:26	which um
0:16:27	which is
0:16:28	yeah
0:16:28	thing that uh that uh after kenny had
0:16:31	um and so
0:16:32	really
0:16:33	all we do in implementing the as non here is
0:16:36	two
0:16:36	define a new set of impostors which is simply the union
0:16:39	this is the non impostors in the teen on impulse
0:16:42	and
0:16:42	we get is the new scoring function
0:16:45	um
0:16:46	that
0:16:47	that looks
0:16:47	pretty similar to any other normalisation function that we have except we just simply
0:16:52	the two
0:16:53	uh added to normalise
0:16:55	scores and this becomes R S nine
0:16:57	um what the the first time there
0:16:59	refers to
0:17:00	um use of W S which is uh
0:17:03	which is the estimate parameters associated with
0:17:05	on your your model
0:17:07	on W S
0:17:07	and then you have your score
0:17:09	um are you have you mute
0:17:10	so the uh sixty a
0:17:12	which
0:17:13	becomes simply that's not
0:17:15	you know
0:17:16	your uh
0:17:17	the
0:17:19	the test utterance
0:17:19	so
0:17:20	so what this gives us now is universal procedure for exactly normalisation parameters and
0:17:25	correspondingly simple method
0:17:26	for score normal
0:17:28	and then
0:17:29	the the the last step of um
0:17:31	normalisation that we we
0:17:33	i explored here
0:17:34	was
0:17:35	um previously discussed in the genes discussion
0:17:37	but um that's sort of what it looks like
0:17:39	um i just as a quick reminder
0:17:42	so
0:17:42	we have now are are some maybe some some experiments that were right
0:17:46	and the system that i used here was
0:17:48	really the same system that um
0:17:50	that niche in previously had
0:17:51	um
0:17:53	given that we're working together so
0:17:54	yeah you're you're standard
0:17:56	parameters for for the system setup
0:17:58	and
0:18:00	and then the
0:18:01	the table of the corpora that we use
0:18:03	um
0:18:04	which
0:18:05	can take a more detailed look at at some other
0:18:07	time like
0:18:08	but
0:18:09	um that the idea of with with this
0:18:11	the protocol of our experiments was
0:18:13	we ran
0:18:13	we wanted to test
0:18:15	our results are based on the female part of the two thousand eight
0:18:18	nist sre
0:18:19	on data set and we fix
0:18:21	our decision an adaptation threshold data
0:18:24	um as the optimal
0:18:26	min dcf
0:18:27	a posteriori
0:18:28	decision threshold and development data uh
0:18:30	nist two thousand
0:18:34	um and so these are like
0:18:36	the the ten second condition results what we
0:18:38	we cared mostly about the ten second
0:18:40	results here
0:18:41	and um
0:18:43	and these are these are the results that we see we we can see that uh
0:18:46	in terms of the minimising detection cost function the
0:18:49	the adaptation based
0:18:50	ct norm on
0:18:51	achieves the best uh min dcf
0:18:54	um whereas
0:18:55	and there
0:18:56	the
0:18:57	then normalised cosine
0:18:59	on that
0:19:00	then is in previously discussed did it also very well
0:19:03	um
0:19:04	did very well
0:19:04	yeah um
0:19:05	or
0:19:06	for everything else
0:19:07	um the english trials for equal error rates
0:19:10	and
0:19:10	also than than in dcf or
0:19:12	all all trials
0:19:17	and but at the same time we we can also notice
0:19:20	that uh
0:19:20	as normal also come uh
0:19:22	achieves
0:19:23	uh good results very competitive
0:19:25	um
0:19:26	and in in in some cases even even better
0:19:29	uh then
0:19:30	then then then Z T norm
0:19:32	and this
0:19:33	um at least for english
0:19:34	whatever
0:19:35	so
0:19:36	uh
0:19:38	so
0:19:39	in order to validate our results we also tried
0:19:41	our work
0:19:41	in the in the i guess this
0:19:44	uh the
0:19:45	log of the conversation the longer
0:19:47	utterances and um
0:19:49	and and this one in this case uh regions
0:19:51	normalisations
0:19:52	stan actually sweat
0:19:53	our results across the board
0:19:55	um
0:19:55	in
0:19:56	in achieving that the best results here
0:20:00	um but
0:20:00	and at the same time there there are a couple things
0:20:03	to to take no
0:20:04	um
0:20:06	in in that um
0:20:07	well we can see here
0:20:09	is that uh
0:20:11	what are our proposed adaptation algorithm is
0:20:13	um
0:20:14	successful
0:20:15	in
0:20:16	improving performance
0:20:17	results uh regardless of the normalisation procedure
0:20:20	um we
0:20:21	it this is obviously consisting with the notion that unsupervised adaptation with of course an appropriately problem
0:20:26	chosen threshold
0:20:27	on should be at least as good as
0:20:29	the hopefully better than on the baseline method without
0:20:32	adaptation
0:20:33	um
0:20:34	and
0:20:35	the next thing is that we have our simplified as an approach
0:20:38	for
0:20:38	performance
0:20:39	competitively with the more complicated traditional teaching on but
0:20:43	um
0:20:44	and of course
0:20:45	um what we've seen is that
0:20:46	the best result is ultimately obtained using cosine normalisation
0:20:50	um
0:20:50	and as a result we i i think
0:20:53	one of the
0:20:53	cooler things is that we we we seem to come full circle with the study
0:20:57	and the story of a score normalisation techniques in that
0:21:00	um
0:21:02	in the beginning we needed normalisation techniques in order to better
0:21:05	sort of calibrate our scores on for fixed decision threshold
0:21:08	and then
0:21:09	however as we got to like the most complicated ct norm
0:21:11	we actually started going backwards and sort of simplifying things
0:21:14	um into an as normal
0:21:16	which is much easier to calculate
0:21:18	and then
0:21:18	and then now it's
0:21:19	it's almost a
0:21:20	um
0:21:22	it that the parameters that we need
0:21:24	are
0:21:25	are not speaker dependent at all there is no
0:21:27	need in the in the normalised cosine distance to actually
0:21:30	have
0:21:31	um
0:21:32	each speaker parameter
0:21:33	uh each uh that the parameters of each distribution calculated
0:21:37	um for each speaker it's just it's pretty it's a pretty universal
0:21:40	um
0:21:41	set of
0:21:42	parameters that need you
0:21:44	um and
0:21:45	and now um i there's a
0:21:47	bit of
0:21:47	work
0:21:48	that i i i
0:21:49	brought up earlier that where we decided that a lot
0:21:52	maybe
0:21:53	um there's a better way to improve
0:21:55	um our our score combination function and this is
0:21:58	um
0:21:59	basic some basic ideas i don't want to go over
0:22:01	that we're currently working but we don't have any uh
0:22:04	really we don't have any significant improvement
0:22:06	in our results just yet
0:22:08	in terms of uh
0:22:10	in terms of our results
0:22:11	um however the idea is weighted averaging because
0:22:14	are currently proposed method for combining scores really treat
0:22:17	every vector in the set
0:22:18	oh so you control factors
0:22:20	as equally important however
0:22:22	at the end of the day on the only
0:22:25	back to that unequivocally
0:22:26	um
0:22:27	the long
0:22:28	speaker
0:22:28	to the speaker as is the initial one long vector as such
0:22:31	maybe it makes more sense to wait that vector is a little bit higher
0:22:35	um then
0:22:36	then we have uh then then the rest of the
0:22:39	training utterances that that we admit
0:22:40	because
0:22:41	the presence of false alarm adaptation updates
0:22:43	um in which
0:22:44	it has utterances incorrectly admitted
0:22:46	um will have an adverse effect on all sorts
0:22:49	test
0:22:49	i said maybe
0:22:50	R R score combining functions should take
0:22:52	take the following into one
0:22:54	into account
0:22:55	um
0:22:56	where
0:22:57	uh where we we each score by
0:22:59	a coefficient eight
0:23:00	where the school is a is um
0:23:02	is the unnormalised score like the something like the cosine similarity
0:23:06	on that ranges
0:23:07	between negative one and one
0:23:08	so it can be seen as a way
0:23:10	teach score
0:23:11	and then that that's sort of
0:23:13	well we can look at
0:23:14	on an a quick visualisation and all morning at time
0:23:17	right now
0:23:17	is that uh we can we can simply
0:23:20	take a look at it this way where
0:23:21	or if you're
0:23:22	initial identity speaker vector is uh
0:23:25	W someone is on most important for the next you vectors might be erroneous because
0:23:29	uh based on
0:23:30	so you're threshold you're only allowing
0:23:32	um the green circle the region in the green circle
0:23:34	in in so
0:23:35	um if you're to speaker identity is S
0:23:38	then
0:23:38	you may incorrectly
0:23:40	uh allow some factors in
0:23:42	um however
0:23:43	as you get
0:23:45	hmmm
0:23:46	um
0:23:47	yeah
0:23:48	a as you get more and more
0:23:49	vectors then we can we can be more more certain that they belong
0:23:53	into the speaker's identity
0:23:54	and so
0:23:56	what we would have
0:23:57	is like
0:23:57	you add a training vector and use it that
0:23:59	space little bit and then
0:24:01	and then maybe you you add in an incorrect one
0:24:04	um but then you i don't more correct one and and
0:24:06	uh as a result after
0:24:08	after a couple more these utterances you can see
0:24:10	finally that maybe
0:24:11	you get you get the right one however you did that false alarm one maybe you should
0:24:15	take it out or so
0:24:16	like that
0:24:17	um
0:24:17	and so
0:24:18	that's sort of like what we're looking at working on in in feature and and we're still looking to improve
0:24:22	the score combination function
0:24:24	welcome to any ideas and
0:24:25	and um
0:24:27	and one of the ideas
0:24:28	also is that uh
0:24:29	though it's not allowed in in this protocol say but
0:24:32	but since we're most airport in the beginning after a while
0:24:35	maybe we can go take another look
0:24:36	at the training vectors and correct any errors from the beginning
0:24:39	uh
0:24:41	and this is easy to do because we don't actually modify
0:24:43	modify vectors at all
0:24:45	um
0:24:47	and so
0:24:48	we have a final summary is that we have
0:24:51	uh propose they
0:24:52	uh a method for unsupervised speaker adaptation um in the use of total variability cosine scoring
0:24:58	um
0:24:59	we have simple and efficient method for score combination
0:25:02	and the fixed a priori decision threshold before class
0:25:05	and um and this
0:25:06	method here can also easily comedy
0:25:09	all score normalisation procedures
0:25:11	um
0:25:12	and
0:25:12	with respect to score normalisation
0:25:14	um discuss some
0:25:16	some of the more
0:25:17	then you were
0:25:18	non C T norm
0:25:19	um ideas which are like a snore
0:25:21	and the normalised cosine that
0:25:23	that the gene
0:25:24	um
0:25:24	a talk about it
0:25:26	so
0:25:27	oh thanks
0:25:37	uh the the
0:25:38	questions for uh
0:25:39	steven
0:25:48	you different paper
0:25:49	the last one was proposed by
0:25:52	each student was in the
0:25:54	we look at this problem and trying to sure but
0:25:57	if you are using a fixed threshold
0:26:01	what
0:26:01	never
0:26:02	the eighteen you uh speaker model
0:26:05	you want thinking about the new uh
0:26:08	cost function proposed by needs to be sure
0:26:11	you know
0:26:11	why
0:26:12	new terms
0:26:13	to uh
0:26:14	you don't have the will to the beginning
0:26:16	to have at least we all uh
0:26:20	we need if we should all over the news for sure
0:26:23	and we propose in this paper
0:26:25	the way to do that they should like
0:26:27	you are doing to schools
0:26:28	we continue so that they shouldn't
0:26:30	oh
0:26:31	just
0:26:32	try to evaluate the confidence so
0:26:34	which we all
0:26:35	and it would look easy
0:26:37	wait
0:26:38	this real with
0:26:39	the confidence
0:26:41	and
0:26:41	use it for good reading put increasing between you do twice
0:26:44	fig legal
0:26:47	um
0:26:48	so so the question is
0:26:49	whether or not i tried
0:26:50	using
0:26:52	we should could be first
0:26:53	you using
0:26:54	but
0:26:55	using a six
0:26:57	sure
0:26:58	two
0:26:58	so if you well
0:27:00	using or not to be only two two
0:27:03	to view speaker model
0:27:05	using
0:27:06	good solution but
0:27:08	would be
0:27:09	we use to
0:27:11	um
0:27:12	if um
0:27:12	upon hearing the question correctly um
0:27:16	i think so yeah uh that the use of a a fixed
0:27:19	decision to
0:27:20	four
0:27:20	i i believe would have
0:27:22	um
0:27:22	well i
0:27:24	at the end of the i think it makes things simple
0:27:26	um
0:27:28	as opposed like
0:27:29	and then
0:27:30	our
0:27:32	using a fixed decision threshold
0:27:34	but at the same time using um a
0:27:36	of varying score
0:27:37	combination function that
0:27:39	that will
0:27:40	wait
0:27:41	wait
0:27:41	scores
0:27:42	um
0:27:43	oh
0:27:43	each
0:27:45	of each training
0:27:46	training utterance that we have
0:27:47	um i i think that that's like a pretty way good way to
0:27:51	pretty good
0:27:51	good way to do it but uh i guess
0:27:53	we we could
0:27:54	talk more about it
0:27:55	later yeah i
0:27:56	i would just be so simple but we question
0:27:58	yeah
0:27:59	when you
0:28:00	compute you
0:28:01	try to imagine you compute to weighted school board would exist
0:28:05	resort materials
0:28:07	okay
0:28:08	i'm just the blah you addition reachable
0:28:11	on this school
0:28:12	you think it would be to work with them until now
0:28:17	um
0:28:19	oh computing the score with
0:28:20	all the
0:28:22	every single
0:28:23	sure trial
0:28:25	so i'm i'm not
0:28:28	okay
0:28:30	uh jumps also i i also have the non certainly we just will on it
0:28:35	uh i also have an onset to uh
0:28:38	your question about
0:28:40	prior
0:28:41	uh
0:28:42	and
0:28:42	uh
0:28:43	i have a a rather mixed in the comment
0:28:46	on on this in my presentation tomorrow
0:28:49	so uh
0:28:50	um
0:28:52	it's
0:28:52	this
0:28:53	this
0:28:53	do that
0:28:56	might be a last question too
0:28:58	steven
0:29:03	okay
0:29:03	it's time to speak again

Unsupervised Speaker Adaptation based on the Cosine Similarity for Text-Independent Speaker Verification

SESSION 4: Speaker and language recognition – scoring, confidences and calibration

Přidáno: 14. 7. 2010 11:08, Autor: Stephen Shum, Najim Dehak (Massachusetts Institute of Technology), Reda Dehak (Laboratoire de Recherche et de Developpement de l'EPITA), James Glass (Massachusetts Institute of Technology), Délka: 0:29:13