Přepis řeči - INTRA-SESSION VARIABILITY COMPENSATION AND A HYPOTHESIS GENERATION AND SELECTION STRATEGY FOR SPEAKER SEGMENTATION

0:00:13	a segmentation
0:00:16	so a a good are known and thank you for coming
0:00:20	yeah and present now where we have do not the university of that was i needs so on
0:00:25	this is a variability ability compensation and for the segment eight for speaker segmentation and to speak that phone conversation
0:00:32	and we are also present in our
0:00:34	a a technique to a it several hypotheses hypothesis and
0:00:39	a for a given recording and to select the that best hypothesis
0:00:42	there a segmentation the
0:00:44	and so uh
0:00:46	these work is focused on a the segmentation of to speaker conversation so it's
0:00:52	is a speaker diarization problem
0:00:55	a a a a a a so we we i'm at the answering the question was but one
0:00:59	but
0:01:00	and
0:01:01	um and it's an is your task seemed since we and number of speaker is known
0:01:07	and and limited to two
0:01:09	so in this case now win the boundaries
0:01:11	the speaker of that is of the that decision problem so it
0:01:14	we can
0:01:15	can um
0:01:18	we could it as a a a a segmentation problem
0:01:22	so i only there's mean a that in the field of the speaker verification and this has motivated and you
0:01:28	approach just for this some addition of to speaker conversation
0:01:31	yeah many base some factor analysis use eigenvoice voices
0:01:35	and in this approach is uh the um
0:01:37	the
0:01:38	the speaker I D model a a bit a gmm supervector
0:01:42	and that can be a present but this small set
0:01:44	oh or a a a a has a small they mention vector that we use a record the speaker factors
0:01:50	was i mention is much lower that they do you not a gmm supervector more
0:01:55	so a a a the main idea is that
0:01:58	a a a using a a a a a compact speaker a presentation we can estimate the a the parameters
0:02:03	of these presentation
0:02:04	what is more segments
0:02:06	and that's what we do for
0:02:08	for
0:02:09	speaker segmentation
0:02:11	so we eh
0:02:12	a start i stream of speaker factors over the input signal
0:02:17	so we are over or a one second window and
0:02:20	frame by frame we start uh
0:02:23	uh a set of
0:02:24	a a a a a sequence of speaker factor of but that's
0:02:26	and then we cluster these the speaker factors and the two clusters
0:02:30	using pca A plus k-means clustering
0:02:33	a once we have the two class we find a single got a single full covariance
0:02:38	a cost four four
0:02:40	it's just speaker or and we wanna be that we segmentation obtain a first
0:02:43	a a some addition up with and then we would find this some addition using now
0:02:47	some uh for segmentation step
0:02:50	um
0:02:52	using them to mfccs features and the mm a speaker models
0:02:57	and
0:02:58	so the
0:02:59	the but the main contribution of this work is T the
0:03:04	the the in general
0:03:05	uh the of for in on whatever the T
0:03:08	so first
0:03:09	that
0:03:09	i about that that's of but a do we can find a the set so
0:03:14	and and only when a if we have a similar or similar recordings containing in different speakers
0:03:21	we analyse the
0:03:22	by a T percent on these recordings
0:03:25	a a we can fight that there's the by ability this actually this by really the and it's mainly you
0:03:31	to
0:03:32	as because percentage recording so you use you referred to as well with yes and the speaker what every
0:03:37	but
0:03:38	we analyse a a set of a
0:03:40	the accordance don't to the same speaker
0:03:43	a so we can see that there's also whatever lead them on this record
0:03:47	and use usually do you to a
0:03:49	two aspects like the channel or
0:03:52	the most chance of the speaker
0:03:54	and this
0:03:55	this but every is usually um
0:03:57	no known as
0:03:59	and system whatever
0:04:01	but in in addition if we if we analyse
0:04:04	a a recording contained a single speaker and with that many in smaller
0:04:09	so i say
0:04:10	and that we
0:04:11	and a light but in really the this just slices
0:04:14	a a we have see that there is also whatever it with you now recording
0:04:18	and
0:04:19	this but i bit is usually do to a phonetic balance or
0:04:22	or they
0:04:23	they won't young on the one channel of the of the
0:04:26	of the recording
0:04:28	and we will refer to as by big guess
0:04:30	and that's system whatever you
0:04:33	no approach for speaker segmentation we are only a
0:04:37	but we we are only modeling in speaker whatever every
0:04:41	so the question is a a are you the types of whatever the inter or interest
0:04:45	my ability a fact some of the in performance
0:04:49	well
0:04:50	and
0:04:52	so we just one do we need to compensate for inter system whatever they don't this is somewhat of a
0:04:57	you
0:04:57	we note that in the system whatever decomposition is but important for speaker recognition
0:05:02	uh yeah but
0:05:03	well i was able to say that it's not
0:05:05	i so be important for speaker some shown but the
0:05:09	is presentation yes so that
0:05:11	channel factors it you helps us well
0:05:13	so i i i have some preliminary experiments that P
0:05:17	then
0:05:18	the help but
0:05:19	may keep my
0:05:20	but we believe that it's it should and had so much because you not that is decision that's you don't
0:05:25	see the same a speaker over different sessions
0:05:27	so it's i was the same as the the speakers are
0:05:30	a single session
0:05:31	you don't have a higher information of the speakers
0:05:34	actually we we believe that it does is a whatever ready to make up the body the speakers and diarization
0:05:38	task
0:05:39	because the channel is information that can
0:05:41	how do you to separate the speak
0:05:44	and what about the some whatever it
0:05:46	so what what usually in the feel of speaker recognition a a state-of-the-art system doesn't take only a a a
0:05:53	a that that's system and don't take a
0:05:55	and take into account intersession variability
0:05:57	seems yeah um they used a whole conversation to to be a model
0:06:02	uh but
0:06:03	we we think it's
0:06:04	but important for speaker some of this and their efficient because um many
0:06:09	and the state of the systems are based on
0:06:11	the clustering of various mark be or segment
0:06:14	and
0:06:15	we can compensate do but i believe them and be segments for a given a speaker
0:06:19	okay
0:06:19	clustering process should be seen
0:06:23	so that's
0:06:24	but we try to do so you been a
0:06:27	a each dataset contains several speakers sensor out according to per speaker
0:06:31	we kind of thing a team of speaker factors for from each recording
0:06:35	and then
0:06:36	a a we can see that every such and as a different class
0:06:40	and we
0:06:41	more more to the speaker and that's system some whatever the guess between class body on seems we believe that
0:06:45	more in a speaker and that's as them whatever the have
0:06:49	to two separate speakers
0:06:50	and in a recording
0:06:52	and we model the session and whatever you S within class but yeah
0:06:56	so we we this framework in
0:06:59	it's C C to to apply a one known techniques has
0:07:02	linear discriminant analysis of much might you can class but as to minimize we class variance
0:07:07	are also within class covariance normalization
0:07:09	a a to normalize the variance
0:07:12	of
0:07:12	for every class
0:07:13	so it's
0:07:14	the identity map fix
0:07:16	so this to thing this have been successfully applied the uh
0:07:20	for intersession compensation and in a speaker recognition
0:07:24	they
0:07:24	the in i of sister
0:07:27	so to evaluate this this of these two approaches we use a a a a a a nice are we
0:07:32	weight
0:07:32	a a a some channel condition containing more than two thousand five minute telephone conversations
0:07:38	i and that the speech nonspeech a what's are given
0:07:41	and we miss the performance and that's of the speaker segmentation or or or speaker or
0:07:46	a a a a a part of the that is as an hour rate so a as we
0:07:49	we have some parts that the speech nonspeech segmentation and "'em" we don't take into account overlap speech
0:07:54	the that is as an hour rate is the same as T segmentation or or or speaker are
0:07:59	and uh a C us use what we you we assume a don't twenty five second people score
0:08:06	and here we have the the results
0:08:08	for the system using in a small ubm
0:08:10	a a two hundred fifty six gaussians
0:08:12	a prior
0:08:13	and mfccs features
0:08:16	and in this case we we don't we don't use that the segmentation of steps
0:08:20	a a a a we can see that our get in two percent some or or
0:08:25	um using intersession session variability compensation and W C C and we
0:08:28	twenty a speaker factors
0:08:30	a a the sum an or is that used to a two point five
0:08:34	you also can see that another other another baseline
0:08:36	a a with fifty speaker factors that it
0:08:39	a slightly better
0:08:40	know
0:08:40	not much but slightly better
0:08:42	and to to try a L D A for them dimensional direction
0:08:47	and we can see that the L use had been
0:08:49	a a a a but i an i and W C C N any
0:08:52	is better
0:08:53	the obtained a two percent a a of segmentation or
0:08:57	and even the combination of what you are is is not better than
0:09:01	now
0:09:02	a and directly W C N
0:09:05	but uh we try with this
0:09:07	systems uh a after the resegmentation and it was surprising pricing that
0:09:11	there was some of the step
0:09:14	um
0:09:15	make more or less equal that was also
0:09:17	using twenty you're fifty speaker factors so
0:09:20	the inter just but every decomposition with W Z and just your working
0:09:24	and
0:09:24	giving the a an improvement
0:09:27	but it seems that it's a useful to improve the number of speaker factors so
0:09:32	we were a little disappointed with
0:09:33	this
0:09:34	because with all it to suit help
0:09:36	and we that i know where use but it meant to that not in the paper
0:09:41	representing here
0:09:43	a without a target ubm M um more features
0:09:46	so in this case increase in the number of is because fact of his help
0:09:49	so we can see that in this case or baseline E we the speaker factors
0:09:54	thanks so one point eight
0:09:55	segmentation or which is a a lower what before he was
0:09:58	two point one
0:10:01	and now that we use channel again but use the your work to one point four
0:10:05	in in in this case a a in this case we we also increase the number of speaker factors and
0:10:11	to test the L A
0:10:12	and we see that the the eighties
0:10:14	yeah is had been you've and
0:10:16	a a more than before
0:10:18	a
0:10:19	and and also that
0:10:20	our but configuration now is used to combine L D A plus W C C N
0:10:25	so it seems that uh it's is it's to question that a
0:10:29	the base and we'd have the speaker factors he's not better than the base them with fifty speaker factors but
0:10:34	that there and the egg
0:10:35	a we can take advantage of how more speaker factors
0:10:40	a our best result is
0:10:41	one point three some our
0:10:45	so on and on the other hand
0:10:47	you we propose you know so i think need to to you know it several segmentation a
0:10:52	hypothesis
0:10:54	and to select the best one a
0:10:56	base of based on a set of from mister
0:10:59	a so what we do is we it adaptively pretty it become the composition to to have this
0:11:06	a a a a a a did we obtain four levels of splitting in as we can see the figure
0:11:10	a a a a a and we segment every um
0:11:12	every a slice with a propose a a system three
0:11:16	then for every level we set at the best the slides this
0:11:19	and we combine them to be able to speaker model
0:11:23	and then we this to speaker models we to segment
0:11:26	the whole recording
0:11:27	using a
0:11:28	i there with some segmentation and mfccs features
0:11:31	you and i speak a speaker model
0:11:34	a to select the what to select the best segment that slices is and the best level um on this
0:11:39	four
0:11:40	we use a a a a a a complete as missiles and also my you're voting stuff
0:11:46	sorry components most of that were using this work were where a bias use information criterion
0:11:51	a a a a a a we using mfccs speech their sign you a speaker models to compute a
0:11:56	big
0:11:57	and and the K yeah these things these dancing the speaker factors space
0:12:02	so we were using gaussian and
0:12:04	a a speaker models and a that space and computing the K U
0:12:07	this stuff between what more
0:12:09	a and
0:12:10	to fuse both compute as were are using the a
0:12:13	a a quite toolkit
0:12:15	a well no for speaker verification
0:12:18	and uh the in the weights of the
0:12:21	a diffuse fusion weights were optimized to separate do for those
0:12:26	a a a a a of time less that one percent someone channel
0:12:30	okay
0:12:31	so
0:12:31	a a kid we have the results
0:12:33	for these uh i but these is you know channel selection
0:12:36	a strategy
0:12:37	we can see when that when we we are not using seen a
0:12:41	inter session variability compensation
0:12:43	a a a a a a what this solution is improving the results which just but our baseline which choose
0:12:48	two point one
0:12:49	and we're getting one point nine with to our started you
0:12:53	and if
0:12:54	we have a an idea
0:12:56	a a coffee that's much some we could select the best
0:13:00	a a level at every time
0:13:02	we could go that to one point one segmentation or
0:13:05	but
0:13:05	the of competence was of our remote idea
0:13:07	at the mall
0:13:08	a
0:13:09	and using system but every to compare that you would we then get
0:13:12	in a significant improvement improvement was
0:13:15	was not the statistic that
0:13:17	that's it's of this signal is significant
0:13:19	a is so we try to my are what in a that the any help
0:13:22	but that we we wanted to to make it what we
0:13:26	it's some set of complete myself to
0:13:28	because the
0:13:30	the possibilities of for are complete and mysteries
0:13:33	of computers was of a high
0:13:35	and
0:13:36	simple uh stuff that you to fuse
0:13:38	a a segmentation hypothesis
0:13:40	so
0:13:42	yeah
0:13:43	we were not really happy with this
0:13:45	was also with try again again with a lot ubm more features
0:13:49	and
0:13:50	and our best racial for intersession variability compensation
0:13:54	and and also a new set of complete missiles
0:13:57	but this
0:13:57	this is not in the paper a new results
0:13:59	but
0:14:00	oh show in two
0:14:02	a a a a a and we we could but use this segmentation addition or or from one point three
0:14:06	to one point two
0:14:07	and one what some additional or one point zero
0:14:10	and if we put select
0:14:11	i always the best level that we could read used to get two point seven
0:14:15	the channel or
0:14:16	which is
0:14:18	but good use so compared to the
0:14:20	to
0:14:21	based
0:14:21	one
0:14:22	well one point
0:14:25	so
0:14:26	ask completion sort of this work we we have presented to to make those for it a somewhat every to
0:14:31	compensation
0:14:32	we have some that they have for speaker segmentation
0:14:36	a a a a a a change that W C N of things better performance than that done of the
0:14:40	eight
0:14:40	and it's
0:14:41	some somehow similar to of the a plus but C C N
0:14:45	that
0:14:45	but similar to the combination of both
0:14:48	um
0:14:49	in the number of a speaker factors increase greece of the computational cost
0:14:53	a it seems that W C N it's
0:14:56	but there's that the word for
0:14:57	a
0:14:59	should should word for low computational cost applications
0:15:02	but the of course a for our best computer and computational cost is not a problem all was computation is
0:15:08	using
0:15:08	a high number of speaker factors and all the egg it uses here
0:15:12	and
0:15:14	we we we have a summary of the results yes we might they our but so that this is one
0:15:19	point three
0:15:20	i
0:15:20	we the a system where a the from one point nine to one point three
0:15:25	and
0:15:26	and also
0:15:27	a a a note that
0:15:28	probably that but used in is had been a lot because
0:15:31	a a a a a a a close or or or because of for but that in the study so
0:15:36	or in is that that use you seen pca plus k-means means
0:15:39	as initialization
0:15:41	so not i seen the the
0:15:44	he within class covariance for a be a speaker is probably had been the K means that assumes that they
0:15:48	all the
0:15:49	and class is have the same class so
0:15:52	at yeah why i are not quite as
0:15:53	a think i
0:15:54	so
0:15:55	so it probably because of the
0:15:57	a
0:15:58	a W C i have some
0:16:01	and also we have present a hypothesis generation and selection technique
0:16:05	which C can prove to this like the results
0:16:08	and for our best configuration we can use D some addition are from one point three to one point to
0:16:13	with a large you
0:16:15	um
0:16:16	think that's all
0:16:17	thank you match
0:16:24	you you we use time for questions
0:16:27	we
0:16:38	yeah
0:16:40	yeah
0:16:43	yeah
0:16:43	i just to one they mention
0:16:46	i didn't mention it because it's
0:16:47	it's in another
0:16:49	but but so this is much more to produce on
0:16:52	on really ability but
0:16:54	yeah P C A
0:16:56	okay we keep just one mention
0:16:59	a a a to initialize a a and then became means we use all that mentions but we need to
0:17:03	like this
0:17:04	the a means of a k-means means we
0:17:07	pca C a of show
0:17:17	no
0:17:18	no
0:17:22	i
0:17:29	yeah
0:17:38	hmmm
0:17:40	yeah
0:17:46	well
0:17:48	yeah yeah
0:17:49	sure
0:17:50	yeah but uh i i mean
0:17:52	now experiments i i am i keeping one i'm and C N N
0:17:55	uh
0:17:56	maybe
0:17:57	you know
0:17:58	is not the best you can do but
0:18:00	to one they mention it's not
0:18:01	but it was so in is is usually the first dimension
0:18:05	of
0:18:06	of the pca put these is the best want to to it this because but it still i we are
0:18:10	getting a about eighteen percent that is is an error right
0:18:13	just using one dimension
0:18:14	so we are not sure that i
0:18:17	the best presentation
0:18:23	yeah
0:18:30	hmmm
0:18:31	so you we we were try to
0:18:33	just plug yeah C A output
0:18:36	be more they mentioned for you and all all the images to the came
0:18:41	the are questions
0:18:47	and let's thing the speed than then one the speech was reduced

INTRA-SESSION VARIABILITY COMPENSATION AND A HYPOTHESIS GENERATION AND SELECTION STRATEGY FOR SPEAKER SEGMENTATION

Miscellaneous Speaker Identification

Přednášející: Carlos Vaquero, Autoři: Carlos Vaquero, Alfonso Ortega, Eduardo Lleida, University of Zaragoza, Spain